2403.09724

# ClaimVer: Explainable Claim-Level Verification and Evidence Attribution of Text Through Knowledge Graphs > Work does not relate to position at Amazon. Abstract In the midst of widespread misinformation and disinformation through social media and the proliferation of AI-generated texts, it has become increasingly difficult for people to validate and trust information they encounter. Many fact-checking approaches and tools have been developed, but they often lack appropriate explainability or granularity to be useful in various contexts. A text validation method that is easy to use, accessible, and can perform fine-grained evidence attribution has become crucial. More importantly, building user trust in such a method requires presenting the rationale behind each prediction, as research shows this significantly influences people’s belief in automated systems. Localizing and bringing users’ attention to the specific problematic content is also paramount, instead of providing simple blanket labels. In this paper, we present ClaimVer, a human-centric framework tailored to meet users’ informational and verification needs by generating rich annotations and thereby reducing cognitive load. Designed to deliver comprehensive evaluations of texts, it highlights each claim, verifies it against a trusted knowledge graph (KG), presents the evidence, and provides succinct, clear explanations for each claim prediction. Finally, our framework introduces an attribution score, enhancing applicability across a wide range of downstream tasks. ClaimVer: Explainable Claim-Level Verification and Evidence Attribution of Text Through Knowledge Graphs Preetam Prabhu Srikar Dammu 1, Himanshu Naidu 1, Mouly Dewan 1, YoungMin Kim 1, Tanya Roosta 2,4, thanks: Work does not relate to position at Amazon., Aman Chadha 3,4, footnotemark: , Chirag Shah 1 1 University of Washington 2 UC Berkeley 3 Stanford University 4 Amazon GenAI <details> <summary>x1.png Details</summary>  ### Visual Description ## Textual Analysis: Misinformation Fact-Checking Framework ### Overview The image presents a structured comparison of two misinformation claims (vaccines/autism and moon landing hoax) with corresponding fact-checks, Wikidata QIDs, and ClaimVer annotations. The framework uses color-coded text highlighting to dissect claims and refutations. ### Components/Axes 1. **Sections**: - **A**: Vaccines and autism prevalence claim - **B**: Moon landing hoax claim 2. **Fact-Check Sources**: - HealthFeedback.org (Section A) - Google Fact Check Tools (Section B) 3. **ClaimVer Annotations**: - Color-coded text (yellow/orange for claim components, red for refutations) - Wikidata QIDs linked to entities/concepts 4. **Refutation Structure**: - **R1-R3**: Logical breakdown of claim components and their validity ### Detailed Analysis #### Section A: Vaccines and Autism - **Claim**: - "Autism used to be 1 in 10,000. Now it's 1 in 50. Now, where it all coming from? Vaccines are doing it." - Highlighted components: 1. Prevalence shift (1/10,000 → 1/50) 2. Causal attribution to vaccines - **Fact-Check (HealthFeedback.org)**: - "Inaccurate: The link between vaccines and autism has already been disproved in several studies." - **ClaimVer Annotations**: - **R1**: "Prevalence of autism is not directly supported or refuted" - **R2**: "Origin of the increase in autism prevalence is not addressed" - **R3**: "Statement that vaccines are causing the increase in autism prevalence is directly contradicted by the triplet [('autism', 'does not have cause', 'vaccine')]" - **QIDs**: - Q134808 (Vaccine: biological preparatory medicine) - Q38404 (Autism: neurodevelopmental condition) #### Section B: Moon Landing Hoax - **Claim**: - "Image shows mismatch between Neil Armstrong’s spacesuit and boot print left on the Moon, therefore Moon landing was a hoax." - Highlighted components: 1. Visual mismatch in spacesuit/boot print 2. Conclusion: Moon landing was a hoax - **Fact-Check (Google AFP)**: - "False: The moon landing was a significant event and an instance of the Apollo 11 mission, which contradicts the claim." - **ClaimVer Annotations**: - **R1**: "Specific claim about the mismatch between the spacesuit and boot print is not directly supported or refuted" - **R2**: "The triplets directly state that the Moon landing was a significant event and an instance of Apollo 11 mission, which contradicts the claim" - **QIDs**: - Q1615 (Neil Armstrong: American astronaut) - Q495307 (Moon landing: Arrival of a spacecraft on the Moon) - Additional QIDs (Q223571, Q190868, etc.) reference Apollo 11 mission details ### Key Observations 1. **Structured Refutation**: - Both sections use QIDs to anchor claims in verifiable entities (e.g., Q38404 for autism). - ClaimVer annotations dissect claims into logical components (e.g., prevalence, causality). 2. **Color-Coded Logic**: - Yellow/orange highlights claim components; red highlights refutations. - Example: "Vaccines are doing it" (red) directly contradicts Q38404's definition. 3. **Wikidata Integration**: - QIDs link claims to authoritative definitions (e.g., Q134808 defines vaccines as "biological preparatory medicine"). - Moon landing QIDs (Q495307, Q1615) emphasize historical consensus. ### Interpretation The framework demonstrates how misinformation can be systematically debunked using: 1. **Entity-Based Fact-Checking**: QIDs ground claims in verifiable knowledge graphs. 2. **Logical Decomposition**: Breaking claims into sub-statements (e.g., prevalence vs. causality) isolates weak points. 3. **Source Credibility**: Fact-checks from HealthFeedback.org and Google AFP leverage peer-reviewed studies and historical records. The absence of numerical data or visualizations suggests the focus is on **semantic analysis** rather than statistical trends. The annotations reveal that while some claim components (e.g., prevalence shifts) lack direct evidence, others (e.g., vaccine causation) are explicitly contradicted by established knowledge. </details> Figure 1: Demonstration of ClaimVer for claim verification and evidence attribution. (A) Text labeled as Inaccurate by HealthFeedback and ClaimVer’s predictions, rationale, and evidence. (B) Text labeled as False by Google Fact Check Tools and ClaimVer’s outputs. Predictions are color-coded (amber: extrapolatory, red: contradictory); $R_{i}$ : rationale; related wiki entities are displayed in boxes. 1 Introduction Misinformation and disinformation are longstanding issues, but the proliferation of AI tools that can generate information on demand has amplified these issues. Tools for fact-checking are not keeping pace with sophisticated text generation techniques. Even when they are effective, they lack appropriate explainability and granularity to be useful to users. Studies have shown that explanations are crucial for users to build trust in AI systems Rechkemmer and Yin (2022); Weitz et al. (2019); Shin (2021). Therefore, there is a need for a novel human-centric approach to text verification that offers usable and sufficiently granular explanations to inform and educate the user. Most fact-checkers, including widely used ones in deployment, issue blanket predictions that can lead to user misunderstandings. For instance, in Figure 1 (A), we observe that HealthFeedback https://healthfeedback.org/, a fact-checker for medical text, indicates that a misleading statement about the increase in Autism is inaccurate. However, there are multiple claims made in that text, which are not addressed by this tool. In fact, research does show that Autism cases have increased, but this is mostly attributed to increased testing Russell et al. (2015). Our method accurately breaks down the text into multiple claims and shows that the specific claim that vaccines are causing autism is indeed incorrect, attributing it to a fact from the Wikidata Vrandečić and Krötzsch (2014). It also provides a clear rationale as to why the first two claims cannot be determined, as there’s no conclusive evidence present in the KG. Such granular predictions, supported by justifications, significantly improve user confidence Rechkemmer and Yin (2022); Weitz et al. (2019); Shin (2021). Similarly, in Figure 1 (B), we notice that Google Fact Check Tools https://toolbox.google.com/factcheck/explorer provides a blanket label for an utterance denying the moon landing. In contrast, ClaimVer identifies the exact text span that can be conclusively proven incorrect and proceeds to provide specific information about the Apollo 11 mission and its crew members to refute the claim. All verified entities present in the text, along with their Wiki IDs and descriptions, are displayed for user reference. Prior research Rashkin et al. (2023); Yue et al. (2023); Thorne et al. (2019); Aly et al. (2021) typically validates text at the paragraph or sentence level without adequately enhancing user awareness by supplying key details such as rationale, match scores, or evidence. A KG-based approach allows for finer granularity, aiding in pinpointing specific inaccuracies like hallucinations in LLM-generated text or false claims in misleading text. Furthermore, if needed, broader-level metrics can be extracted from this detailed attribution. The assumption of one-to-one mapping between input and reference texts, prevalent in previous methods Rashkin et al. (2023); Yue et al. (2023); Thorne et al. (2019); Aly et al. (2021), does not hold if the given text consists of claims that can be mapped to more than one source. In contrast, utilizing a KG, which represents a consolidated body of knowledge, results in a more comprehensive evaluation. While most previous methods may not support scenarios with information spread across various references, querying a KG can yield triplets originally sourced from multiple documents. Additionally, procuring the specific spans of text required to evaluate claims from large text sources that may span several pages presents many challenges. In contrast, a KG captures only the most important relationships as nodes and links, providing a more efficient way to evaluate the claims. 2 Related Work Research on validating text has been ongoing for the past decade, while the concept of evidence attribution has gained increased attention in recent years, following the advent of generative models. Our method integrates fact verification and evidence attribution. In this section, we discuss recent advancements in both domains. 2.1 Fact Verification Fact verification is a task that is closely related to natural language inference (NLI) Conneau et al. (2017); Schick and Schütze (2020), in which given a premise, the task is to verify whether a hypothesis is an entailment, contradiction, or neutral. Similarly, in fact verification, the task is to check if a given text can be supported, refuted, or indeterminable, given a reference text. Recent studies in this domain show that LLMs can achieve high performance, and can be considerably reliable for verification tasks, even though they are prone to hallucations Guan et al. (2023). In Lee et al. (2020), the authors show that the inherent knowledge of LLMs could be used to perform fact verification. Other works Yao et al. (2022); Jiang et al. (2023b) have shown that using external knowledge is helpful for many reasoning-intensive tasks, and report enhanced performance on HotPotQA Yang et al. (2018) and FEVER Thorne et al. (2018). A wide variety of studies have established that LLMs are suitable for fact verification. For example, Dong and Smith (2021) enhanced accuracy of table-based fact verification by incorporating column-level cell rank information into pre-training. In FactScore, authors Min et al. (2023), introduce a new evaluation that breaks a long-form text generated by large language models (LMs) into individual atomic facts and calculates the proportion of these atomic facts that are substantiated by a credible knowledge base. 2.2 Evidence Attribution The distinction between evidence attribution and fact verification lies in the emphasis on identifying a source that can be attributed to the information. This task is becoming increasingly important, as generative models produce useful and impressive outputs, but without a frame of reference to validate them. In Rashkin et al. (2023), the authors present a framework named AIS (Attributable to Identified Sources) that specifies annotation guidelines and underlines the importance of attributing text to an external, verifiable, and independent source. Yue et al. (2023) demonstrate that LLMs can be utilized for automatic evaluation of attribution, operationalizing the guidelines presented in Rashkin et al. (2023). However, both of these works are primarily designed for the question-answering (QA) task. In contrast, our method is not restricted to QA and is designed to work with text in general. Furthermore, while these previous studies focus on sentence or paragraph levels, our approach extends to a more detailed and granular level of analysis. <details> <summary>x2.png Details</summary>  ### Visual Description ## Flowchart: Claim Verification System ### Overview The diagram illustrates a technical pipeline for claim verification using natural language processing (NLP) and knowledge graph (KG) integration. It demonstrates how an input text is processed through preprocessing, knowledge graph analysis, and a fine-tuned language model to generate structured outputs including claim validation, rationale, and confidence scores. ### Components/Axes 1. **Input Text**: - Example: "Steven Tyler has never been a part of the band Aerosmith." 2. **Preprocessing**: - Sub-components: NER (Named Entity Recognition), Coreference Resolution, KG Entity Linking, Compartmentalization. 3. **Knowledge Graph (KG)**: - Visualized as interconnected nodes (entities) and edges (relationships). 4. **KG Triple Retrieval Algorithm**: - Processes KG data to extract relevant triplets. 5. **Finetuned ClaimVer LLM**: - A language model specialized for claim verification. 6. **Outputs**: - Structured results including Claim, Prediction, Relevant Triplets & TMS, Rationale, and Score (KAS). ### Detailed Analysis - **Input Text**: The example claim states "Steven Tyler has never been a part of the band Aerosmith." - **Preprocessing**: - NER identifies entities (e.g., "Steven Tyler," "Aerosmith"). - Coreference resolution links pronouns to entities. - KG Entity Linking maps entities to KG concepts. - Compartmentalization organizes data into structured units. - **KG Triple Retrieval Algorithm**: - Extracts triplets like `[('Aerosmith', 'has part(s)', 'Steven Tyler')]` with confidence score `1.0`. - **Finetuned ClaimVer LLM**: - Combines retrieved triplets with input text to generate outputs. - **Outputs**: - **Claim**: Repeats the input claim. - **Prediction**: "Contradictory" (indicating the claim is false). - **Relevant Triplets & TMS**: Highlights the triplet `[('Aerosmith', 'has part(s)', 'Steven Tyler')]`. - **Rationale**: Explains that the triplet establishes a relationship between Steven Tyler and Aerosmith, refuting the claim. - **Score (KAS)**: `0.047` (low score indicates high confidence in the claim's falsity). ### Key Observations 1. The system identifies a contradiction between the input claim and KG data. 2. The retrieved triplet `[('Aerosmith', 'has part(s)', 'Steven Tyler')]` directly challenges the claim. 3. The low KAS score (`0.047`) suggests strong evidence against the claim's validity. 4. The rationale explicitly links the triplet to the claim's contradiction. ### Interpretation This flowchart represents a hybrid NLP-KG system for fact-checking. The low KAS score (`0.047`) indicates the model is highly confident the claim is false, supported by the KG triplet showing Steven Tyler's association with Aerosmith. The "Contradictory" prediction aligns with the rationale, demonstrating how KG data and language models collaborate to validate claims. The system's design emphasizes transparency by providing both numerical scores and human-readable explanations (rationale). The example highlights the pipeline's ability to detect factual inconsistencies in claims about entity relationships. </details> Figure 2: Flow of operations in the ClaimVer framework. Identified KG entity nodes during preprocessing inform the extraction of relevant triplets by the KG algorithm. Subsequently, these triplets and preprocessed text are then fed to a ClaimVer LLM, fine-tuned to operationalize the objective function. For each claim, the corresponding text span, prediction, relevant triplets, attribution scores, and rationale are generated. 3 Methodology In this section, we present the methodology for retrieving relevant triplets from the KG, fine-tuning LLM to process text at claim-level, verifying claims, tagging evidence for each prediction, and generating a rationale along with an attribution score that reflects the text’s validity. 3.1 Preprocessing Preprocessing involves multiple steps required to make the input text suitable for the subsequent operations. Since the nodes in a KG typically represent entities, performing Named Entity Recognition (NER) is necessary. In our work, we chose Wikidata Vrandečić and Krötzsch (2014) as the KG source; thus, we use an NER module suitable for Wiki entities Gerber (2023). However, the framework is sufficiently generic to support any kind of KG that models information in the form of triplets. As our analysis is performed at the claim level, coreference resolution Lee et al. (2017) becomes a necessary step to form localized claims that are semantically self-contained. If input text exceeds the context length, which depends on design choices, compartmentalization would be required. As a final step in preprocessing, we perform KG entity linking. This step tags all entities in the text that are present in the KG as nodes. 3.2 Relevant Triplets Retrieval Retrieving relevant triplets is a complex problem that has attracted attention from various research communities, and resulted in multiple approaches to address the challenge. While retrieving direct links between two given nodes in a KG is relatively straightforward, identifying complex paths that involve multiple hops is challenging. In our framework, we use Woolnet Gutiérrez and Patricio (2023), a multi-node Breadth-First Search (BFS) algorithm, to retrieve the most relevant triplets for a given claim present in the KG. This BFS algorithm initiates from multiple starting points and, at each step, searches for and processes all adjacent neighbors before advancing. It constructs a subgraph of visited nodes, tracking their origins, and distances from each BFS’s start. The algorithm expands each search tree one node at a time until paths intersect or reach a predefined maximum length. Upon intersection, it assesses if the discovered path meets the length criteria. If so, it logs the route, utilizing backtracking to trace the path to its origins, while ensuring there are no repetitions or cycles, thus maintaining a connection to a starting node. In our experiments, we allow for a maximum of three hops between any two given nodes, and a maximum of four potential paths. Adopting less stringent conditions leads to less relevant triplets. 3.3 Objective Function Previous works on evidence attribution tasks have established definitions for the categorization of input text with reference to a supporting source Rashkin et al. (2023); Gao et al. (2023); Bohnet et al. (2022); Yue et al. (2023). Similar to the formulation in Yue et al. (2023), we use three categories: Attributable, Extrapolatory, and Contradictory. However, there are two main differences that distinguish our approach from previous methods. First, we verify the input text against facts present in a KG, an aggregated information source constructed by integrating numerous data sources into a structure of triplets, instead of relying on a single reference. This approach eliminates the one-to-one dependency between the text and its information source. Second, we perform attribution with finer granularity, specifically at the claim level, involving a subtask of decomposing the input text into individual claims. We define our categories as follows: - Attributable: Triplets fully support the claim. - Extrapolatory: Triplets lack sufficient information to evaluate the claim. - Contradictory: Triplets contradict the claim. We formulate the objective function of our task as follows: where: - $input\_text$ : input text containing claim(s). - $ret\_triplets$ : retrieved triplets for the input text. - $claim\_span_{i}$ : $i^{th}$ claim extracted as a substring from $input\_text$ . - $claim\_pred_{i}$ : label predicted for $claim\_span_{i}$ . - $rel\_triplets_{i}$ : relevant subset of $ret\_triplets$ for $claim\_span_{i}$ . - $rationale_{i}$ : justification for $claim\_pred_{i}$ . - $n$ : total number of claims in $input\_text$ . This objective function encompasses two main sub-tasks: 1. Decomposing input text into claims. 1. Generating prediction and corresponding rationale for each claim by identifying relevant supporting triplets. 3.4 Fine-tuning LLMs The objective function shares similarities with the well-studied task of NLI Conneau et al. (2017); Schick and Schütze (2020). LLMs achieve state-of-the-art performance for NLI Chowdhery et al. (2023), making them a suitable choice to operationalize the objective function. Additionally, Yue et al. (2023) shows that LLMs can be used to automatically evaluate attribution to a given information source. However, these prior methods do not involve a complex sub-task, which is central to the proposed objective function, i.e., decomposing the input text into text spans that correspond to separate claims in the presence of multiple claims. It is crucial to perform both claim decomposition and attribution for all claims in a single step, as processing each claim individually can lead to an exponential increase in LLM queries, leading to significantly higher computational costs and latency issues. In order to perform attribution at the claim level, we need to fine-tune LLMs specifically for the proposed objective function (see § 3.3) using a custom dataset. This is necessary because, as of this writing, even the state-of-the-art model, OpenAI’s GPT-4 Achiam et al. (2023), does not perform satisfactorily right out of the box. Our custom dataset, built using two sequential complex prompts with GPT-4, enables us to fine-tune significantly smaller models. This approach distills the performance of a large proprietary model using a multi-query prompt pipeline into small open-source models with a compact zero-shot prompt. We make the weights of the fine-tuned models publicly available weights available on HuggingFace. We selected eight open-source LLMs with diverse sizes, ranging from 2B parameters to 10B parameters, to perform the fine-tuning: Gemma-2B-IT-Chat Team et al. (2024), Phi-3-mini-4k-Chat Javaheripi et al. (2023), Zephyr-7B-Beta-Chat Tunstall et al. (2023), Mistral-7B-v0.3-Chat Jiang et al. (2023a), Llama3-8B-Chat Touvron et al. (2023), and Solar-10.7B-Chat Kim et al. (2023). The models were fine-tuned using LoRA Hu et al. (2021) with 4-bit quantization and adapters with rank 8 Dettmers et al. (2024). The context length was set to 4096 tokens (for additional training details, refer § A.1) All models converged after 2 epochs, and high ROUGE-L Lin (2004) scores greater than 0.658 were achieved for each model. The instruction prompt used for fine-tuning is presented in Figure 3. Analyze text against provided triplets, classifying claims as "Attributable", "Contradictory", or "Extrapolatory". Justify your classification using the following structure: - "text_span": Text under evaluation. - "prediction": Category of the text (Attributable / Contradictory / Extrapolatory). - "triplets": Relevant triplets (if any, else "NA"). - "rationale": Reason for classification. For multiple claims, number each component (e.g., "text_span1", "prediction1",..). Use "NA" for inapplicable keys. Example: "text_span1": "Specific claim", "prediction1": "Attributable/Contradictory/Extrapolatory", "triplets1": "Relevant triplets", "rationale1": "Prediction justification", ... Input for analysis: -Text: {Input Text} -Triplets: {Retrieved Triplets} Figure 3: Instruction prompt for fine-tuned LLMs. 3.5 Computing Attribution Scores For various downstream tasks, such as ranking and filtering, a continuous score that reflects the validity of a given piece of text with respect to a KG is desirable. We propose the KG Attribution Score (KAS), which accomplishes this task with a high level of granularity, and is detailed in this section. 3.5.1 Claim Scores where, $y_{i}$ is $claim\_pred_{i}$ . For each claim, we assign a score that reflects the level of its validity, ranging from -1 (contradictory) to 2 (attributable). If a claim is predicted to be extrapolatory, yet has one or more relevant triplets, we assign that claim a score of 1, as there is still relevant information available even though it may not be sufficient to completely support or refute the claim. However, if there are no triplets at all, along with an extrapolatory prediction, we assign 0 as it does not add any useful information. While decomposing claims, the model might occasionally omit words, typically stop-words, and we assign 0 in those cases as well. 3.5.2 Triplets Match Score (TMS) This score reflects the extent of the match between the relevant triplets and the corresponding claim, and it can also serve as a proxy for the prediction confidence. Even though the prediction is made at the claim level, the triplets match score considers word-level matches in the computation. It can be computed as follows: where, $E(claim\_span_{i})$ and $E(rel\_triplet_{i})$ represent the sets of entities in $claim\_span_{i}$ and $rel\_triplet_{i}$ , respectively. $SS$ is the semantic similarity computed using the cosine similarity of text embeddings, and $EPR$ represents the ratio of entities in $E(claim\_span_{i})$ that are also present in $E(rel\_triplet_{i})$ . The parameters $\alpha$ and $\beta$ can be adjusted as needed; in our experiments, we use 0.5 for both. In cases where examples of an entity retrieved from the KG are used to support the prediction, instead of the entity itself, we may not have a direct overlap, and thus semantic similarity would be helpful. $EPR$ rewards the direct use of the entity, so a balance between both may be ideal in most cases. 3.5.3 KG Attribution Score (KAS) For the final KG Attribution Score (KAS), a continuous score between 0 and 1 is desirable, as this facilitates various downstream applications such as ranking, fine-tuning, and filtering. This can be achieved using a Sigmoid function. However, the standard Sigmoid function treats positive and negative scores equally. In most cases, higher penalties should be assigned for erroneous text than rewards for valid text. This requirement can be met using a modified Sigmoid function that penalizes mistakes by a factor of $\gamma$ : $$ \displaystyle\sigma_{\text{mod}}(x,\gamma)=\frac{1}{1+e^{-\gamma\cdot x}}, \displaystyle\text{where }\gamma=\begin{cases}\gamma=3&\text{if }x<0,\\ \gamma=1&\text{if }x\geq 0,\end{cases} \tag{4} $$ In our experiments, we set the value of $\gamma$ to 3. Finally, the modified Sigmoid function, applied to the summation of triplet match scores and claim scores, is used to generate KAS: $$ \displaystyle\text{KAS}=\sigma_{mod}(\sum_{i=1}^{n}[ \displaystyle TMS_{i}\cdot cs(y_{i})],\gamma) \tag{5} $$ 4 Dataset | Train Test | 3400 1000 | 5342 1677 | 998 316 | 3546 1068 | 798 293 | | --- | --- | --- | --- | --- | --- | Table 1: Distribution of fine-tuning dataset. Att: Attributable, Ext: Extrapolatory, Con: Contradictory. <details> <summary>x3.png Details</summary>  ### Visual Description ## Table: Claim Analysis with Triplets, Predictions, and KAS Scores ### Overview The table presents six examples of input texts analyzed for claim validity using relevant triplets, prediction scores (TMS), rationales, and KAS (Knowledge Alignment Score) values. Each row demonstrates how triplet relationships and prediction confidence scores contribute to evaluating claims about fictional characters, historical figures, geographical features, corporate activities, religious appointments, and airline operations. ### Components/Axes - **Columns**: 1. **Input Text**: Descriptions of claims with contextual references. 2. **Relevant Triplets**: Semantic relationships extracted from the input text. 3. **Prediction (TMS)**: Confidence scores (0.0–1.0) for triplet relevance, categorized as: - Attributable (directly supports/refutes the claim) - Extrapolatory (indirectly related) - Contradictory (directly opposes the claim) 4. **Rationale**: Explanation of how triplets and predictions support or contradict the claim. 5. **KAS**: Final score (0.0–1.0) indicating overall claim validity. ### Detailed Analysis #### Row 1: George O'Malley - **Input Text**: Fictional character from *Grey's Anatomy* (ABC, USA). - **Triplets**: 1. `("Grey's Anatomy", "characters", "George O'Malley")` 2. `("Grey's Anatomy", "original broadcaster", "American Broadcasting Company")` 3. `("American Broadcasting Company", "country", "United States")` - **Prediction (TMS)**: - Attributable (0.852): Triplets confirm the character and network. - Attributable (0.637): Triplets confirm the network's location. - **Rationale**: Triplets directly support the claim about the character and network. - **KAS**: 0.818 #### Row 2: Robert Swenson as Bane - **Input Text**: Portrayed as a villain in *Batman & Robin* (1997). - **Triplets**: 1. `("Batman & Robin", "cast", "Robert Swenson")` 2. `("Batman & Robin", "director", "Joel Schumacher")` 3. `("Batman & Robin", "director", "Joel Schumacher")` (duplicate) - **Prediction (TMS)**: - Attributable (0.788): Cast triplet supports the claim. - Attributable (0.882): Director triplet supports the claim. - Extrapolatory (0.0): Duplicate triplet adds no new info. - **Rationale**: Triplets confirm Swenson's role and director. - **KAS**: 0.752 #### Row 3: Crater Lake - **Input Text**: Main feature of Crater Lake National Park (deep blue water). - **Triplets**: 1. `("Crater Lake", "located in", "Crater Lake National Park")` (corrected from "Cater Lake") 2. `("Crater Lake", "protected area", "National Park")` - **Prediction (TMS)**: - Attributable (0.942): Triplets confirm location and protected status. - Extrapolatory (0.0): No triplets address water color/clarity. - **Rationale**: Triplets support location but not visual features. - **KAS**: 0.719 #### Row 4: Airbus in Blagnac - **Input Text**: Airbus activity in Blagnac, France (suburb of Toulouse). - **Triplets**: 1. `("Airbus Operations S.A.S.", "country", "France")` 2. `("Airbus Corporate Jets", "headquarters", "Toulouse")` 3. `("Blagnac", "location", "France")` - **Prediction (TMS)**: - Attributable (0.505): Triplets confirm location and corporate presence. - Extrapolatory (0.0): No data on jetliner production. - **Rationale**: Triplets confirm location but not production activity. - **KAS**: 0.583 #### Row 5: Pope Benedict XVI - **Input Text**: Never appointed significant individuals in the Catholic Church. - **Triplets**: 1. `("Rutilio del Riego Jáñez", "appointed by", "Benedict XVI")` 2. `("Rutilio del Riego Jáñez", "religion", "Catholic Church")` 3. `("God", "said to be", "love")` - **Prediction (TMS)**: - Contradictory (0.781): Triplets show Benedict appointed someone. - Extrapolatory (0.065): "God = love" is irrelevant. - **Rationale**: Triplets directly contradict the claim. - **KAS**: 0.248 #### Row 6: Southwest Airlines - **Input Text**: Never operated Boeing 737 models. - **Triplets**: 1. `("Boeing 737 MAX", "operator", "Southwest Airlines")` 2. `("Boeing 737 #1491", "operator", "Southwest Airlines")` - **Prediction (TMS)**: - Contradictory (0.933): Triplets confirm Southwest operated Boeing 737 models. - **Rationale**: Triplets directly refute the claim. - **KAS**: 0.057 ### Key Observations 1. **High KAS Scores** (e.g., 0.818, 0.752) correlate with triplets that directly support/refute claims. 2. **Low KAS Scores** (e.g., 0.057) occur when triplets contradict the input text. 3. **Extrapolatory Triplets** (e.g., "God = love") often add no meaningful context. 4. **Typos** (e.g., "Cater Lake" instead of "Crater Lake") may affect triplet accuracy. ### Interpretation The table demonstrates how triplet-based analysis evaluates claim validity. Directly supporting triplets (e.g., character-network relationships) yield high KAS scores, while contradictory triplets (e.g., papal appointments) lower scores. Extrapolatory triplets (e.g., "God = love") often lack relevance. The analysis highlights the importance of precise triplet extraction and contextual alignment in automated claim verification systems. </details> Table 2: Examples of claim-level attribution by the proposed method. The first column shows the numbered claims in the input text. Second column lists relevant triplets for each claim. Predictions and Triplets Match Score (TMS) are in the third column, while the rationale behind each prediction is in the fourth column. The Knowledge Graph Attribution Score (KAS) is shown in the last column. Model: Solar-10.7B-Chat. Open-domain Question Answering (QA) datasets, such as WikiQA Yang et al. (2015), HotPotQA Yang et al. (2018), PopQA Mallen et al. (2022), and EntityQuestions Sciavolino et al. (2021), as well as Fact Verification datasets like FEVER Thorne et al. (2019), FEVEROUS Aly et al. (2021), TabFacT Chen et al. (2019), and SEM-TAB-FACTS Wang et al. (2021a), provide texts along with corresponding reference contexts or attributable information sources. However, these datasets significantly differ from the type of data required to train and test our proposed objective function, primarily due to two major factors: (i) these datasets predominantly offer samples that are inherently attributable, and (ii) consist of atomic claims and/or one-to-one mappings between input and reference texts. To address the first limitation, prior work Yue et al. (2023) in attribution evaluation introduced new samples by modifying correct answers to generate contradictory instances. Yet, this adjustment alone is not sufficient for our use case because our method requires attribution at the claim level, and necessitates the automatic decomposition input text to claims. Consequently, as this task represents a novel challenge, we developed a new dataset that enables effective training and testing of the objective function. Considering the choice of our KG, which is Wikidata Vrandečić and Krötzsch (2014), we opted for WikiQA Yang et al. (2015) as it is closely associated with the Wiki ecosystem. Given that our method is designed for text validation in general, and is not limited to question answering, we retain only answers and discard the questions. Subsequently, we processed the answers following the steps detailed in Section 3.1, selecting entries containing two or more Wiki entities. This approach resulted in the exclusion of most single-word answers and other responses that are dependent on their corresponding questions and may lack comprehensibility without them. We utilize GPT-4 Achiam et al. (2023) to generate the initial version of the ground truth. Although GPT-4 can adhere to the instructions (refer to Figure 3) to a reasonable degree and responds in the required format with all necessary keys, it still underperforms in the overall task. The most frequent issue observed is the erroneous assignment of prediction labels. To remedy this issue, we designed a detailed prompt tailored for the given task, incorporating techniques such as few-shot, chain-of-thought Kojima et al. (2022), and other strategies OpenAI (2024); Nori et al. (2023) (full prompt in § A Figure 12). We also conducted manual checks to ensure only high-quality samples were retained, as research indicates that high alignment can be achieved with as few as 1,000 samples, provided they are of superior quality Zhou et al. (2023). The final dataset is comprised of two splits: the training split, based on the training split of WikiQA Yang et al. (2015), and a test split, derived from both the test and validation splits. The training split contains 3,400 samples, and since some entries feature multiple claims, there are a total of 5,342 claims within this split. Similarly, the test split includes 1,000 samples and 1,677 claims. The label counts for the claims are tabulated in Table 1. The dataset is publicly shared to facilitate further research in this direction dataset available on HuggingFace. 5 Experiments and Results In this section, we present the evaluation of our claim-level attribution method. The performance metrics of the fine-tuned LLMs, which operationalize the objective function, are presented in Tables 3 and 4. In Table 3, we observe that all models converge and achieve sufficiently high ROUGE-L and ROUGE-1 scores, with Mistral-7B-v0.3-Chat achieving the highest of 0.694 and 0.719 respectively. We also observe that the smaller model, Gemma-2B-IT-Chat with just 2B parameters, is also sufficiently compatible for this task as it attained a decent ROUGE-L score of 0.667. The first task of the proposed objective function (refer § 3.3), decomposing text into multiple claims, is somewhat subjective, and there could be multiple valid approaches due to linguistic complexities. For instance, example 4 in Table 2 has been decomposed into three claims, but the first could arguably be further decomposed to verify whether Blagnac is in France, and whether it is a suburb of Toulouse. Controlling the precise manner of decomposition is challenging, and might necessitate an additional step before the prediction step, involving separate processing for each claim. However, this option could prove to be impractical, as the number of LLM queries could increase exponentially. To accurately compute classification performance, we impose a strict strategy: the text span of the claim, the identified relevant triplets, and the prediction label must all exactly match the ground truth to be considered accurate. In Table 4, the second column indicates number of claims with text spans exactly matching the ground truth responses. Columns 3 to 6 present the accuracy, precision, recall, and F1 scores for these matching claims. The most performant model is Solar-10.7B-Chat, with 1031 exact matches out of 1677 claims in the test set. Additionally, the classification scores in all metrics are above 89%, which clearly demonstrates that the model can reliably differentiate between the classes attributable, extrapolatory, and contradictory. Table 2 showcases the claim-level attribution performed by our method. Each claim in the input text is numbered and color-coded to reflect its prediction: green for attributable, amber for extrapolatory, and red for contradictory. The examples are sorted in descending order by their KAS scores, which reflect the validity of the text. As expected, we observe more green at the top of the table and more amber and eventually red as we move down. Since the Wiki ecosystem is open-domain, we observe that the examples cover a wide range of topics, demonstrating that the method is adaptable to diverse inputs. In the first example in 2, the input text is decomposed into two claims, both of which are attributable. The first claim is supported by a single triplet in the KG, while the second claim can be supported by combining two triplets. The second example presents more challenges for evaluation due to its complex sentence structure, but ClaimVer accurately identifies that the third claim regarding Batman’s Back is neither supported nor refuted by the triplets, as indicated in the rationale. In the third example, we note that the first claim is predicted to be attributable with a high triplet match score of 0.942 since there is a triplet that clearly supports the location description of Crater Lake. However, as there is no information regarding the water characteristics, the second claim is categorized as extrapolatory. In the fourth example, the first claim alone requires three triplets combined as supporting evidence, illustrating the method’s ability to handle complex multi-hop paths within the KG. The second and third claims are predicted to be extrapolatory, since there are no triplets concerning Airbus’s market share, or its activities in Europe, as highlighted in the model’s rationale. It is noteworthy that the context provided in the third claim is crucial for the first claim to be comprehensible, demonstrating why individual claim evaluation may be suboptimal. Interestingly, in the fifth example, the method identifies a specific instance from the KG to refute a general claim, citing the appointment of Rutilio del Riego Jáñez. Similarly, in the sixth example, the method provides specific instances, quoting two distinct Boeing 737 models to demonstrate contradiction with a high triplet match score. | Gemma-2B-IT-Chat Phi-3-mini-4k-Chat Zephyr-7B-Beta-Chat | 2B 4B 7B | 0.667 0.658 0.686 | 0.692 0.685 0.712 | | --- | --- | --- | --- | | Vicuna-7B-v1.5-Chat | 7B | 0.676 | 0.700 | | Mistral-7B-v0.3-Chat | 7B | 0.694 | 0.719 | | Gemma-7B-IT-Chat | 7B | 0.678 | 0.703 | | Llama3-8B-Chat | 8B | 0.679 | 0.705 | | Solar-10.7B-Chat | 10B | 0.689 | 0.714 | Table 3: ROUGE scores on the test set ( $n=1,000$ ). | Gemma-2B-IT-Chat Phi-3-mini-4k-Chat Zephyr-7B-Beta-Chat | 895 882 978 | 77.09 72.22 85.89 | 77.20 78.10 87.41 | 77.09 72.22 85.89 | 74.24 72.86 86.16 | | --- | --- | --- | --- | --- | --- | | Vicuna-7B-v1.5-Chat | 898 | 79.62 | 78.83 | 79.62 | 78.84 | | Mistral-7B-v0.3-Chat | 1002 | 86.63 | 87.03 | 86.63 | 86.73 | | Gemma-7B-IT-Chat | 940 | 82.87 | 84.09 | 82.87 | 83.17 | | Llama3-8B-Chat | 959 | 80.92 | 85.48 | 80.92 | 81.36 | | Solar-10.7B-Chat | 1031 | 89.23 | 89.52 | 89.23 | 89.30 | Table 4: Scores on matching claims in the test set ( $n=1677$ ). #MC: number of matching claims. 6 Discussion The susceptibility of LLMs to generating factually incorrect statements is an alarming concern as LLM-powered services become increasingly popular for seeking advice and information. The democratization of generative models has also had adverse effects, such as increasing misinformation Monteith et al. (2024). To arm end-users with the tools necessary to combat being misinformed, it is crucial to develop text-validation methods that are human-centric, and prioritize user engagement, understanding, and informativeness. We design our method with these principles in mind: we make predictions at the claim level, and identify text spans within the given text, that can be color-coded and presented to the user. The proposed method also generates easily comprehensible explanations along with the prediction and evidence, thus reducing the cognitive burden on the end-user, and making the process user-friendly. The usability and evaluation of these systems should align with human needs and capabilities. Chatbots, such as ChatGPT Achiam et al. (2023), serve a wide array of tasks; therefore, the text validation method should be adaptable to various domains. While KGs like Wikidata Vrandečić and Krötzsch (2014) are considered open-domain, the implementation of more specialized KGs, along with corresponding routing algorithms may be necessary to support a broader range of topics. For instance, a common-sense KG Hwang et al. (2020) would be more useful in validating non-factoid answers that involve logic. Furthermore, the maintenance efficiency of our approach aligns well with the need for sustainable, long-term AI solutions. In a world where information is constantly evolving, the ability to update and maintain AI systems with minimal effort is not just a convenience, but a necessity. This directly ties into the ethical implications of AI, where outdated or incorrect information can lead to harmful decisions. By leveraging existing, well-maintained KGs, we can ensure that AI systems remain accurate and relevant over time. 7 Conclusion In this paper, we present ClaimVer, a framework for text verification and evidence attribution at the claim level by leveraging information present in KGs. In contrast to other methods, ClaimVer eliminates the one-to-one mapping between input and reference text, allowing for layered interpretation and handling of distributed information. In addition to these primary functions, ClaimVer incorporates human-centric design principles by offering clear, concise explanations for each claim prediction—an important characteristic for building user trust and enhancing usability. Furthermore, we introduce an attribution score, which enhances its applicability across a wide range of downstream tasks. Finally, we share ClaimVer fine-tuned LLMs to facilitate further exploration of this research direction. 8 Limitations Limitations of LLMs for Fact Verification. Like most ML models, LLMs are prone to erroneous predictions, which is particularly concerning in sensitive applications such as handling misinformation. Despite this, they remain the most performant techniques for fact verification and related tasks like NLI Yue et al. (2023); Wang et al. (2021b). Therefore, while it is reasonable to use the best option available, fact verification systems relying on LLMs should be utilized with caution and necessary validations. Limitations of Knowledge Graphs. While there are several advantages associated with using KGs, we also acknowledge the presence of known issues, such as knowledge coverage and the efforts required to keep these sources up-to-date. For our solution, we assume that the KG is up-to-date and possesses adequate coverage. However, this may not always be the case, and thus the most suitable technique should be adopted after considering the specific requirements of a particular use case. Another point to consider is that the proposed method does not provide traditional citations to articles, although it may be possible to retrieve that information from the KG, if information source mapping has been properly maintained. Variations in Claim Decomposition. Decomposing text into multiple claims is a complex linguistic task that often results in multiple valid decompositions. Although this may not impact usability if the prediction, rationale, and text spans are comprehensible and supported by facts from the KG, it poses a challenge for evaluating model performance. One potential approach is to operate at the token level instead of the span level, but this would significantly complicate the problem space. Additionally, token-level verification and attribution would require substantially higher compute resources, necessitating further studies to assess their value and impact on system usability and reliability. LLM Reasoning Errors. Previous works have demonstrated that using LLM reasoning for tasks like fact verification, evidence attribution, and NLI can yield impressive results, surpassing alternative approaches Yue et al. (2023); Wang et al. (2021b). However, LLM reasoning can still be flawed. In our work, we impose validations to minimize LLM mistakes by performing membership checks for supporting triplets and string matching for text spans. Yet, validating reasoning remains an open problem with ongoing research efforts. Fine-tuning Dataset Limitations. To build the fine-tuning dataset, we utilized GPT-4 with two detailed sequential prompts designed in accordance with OpenAI’s recommendations OpenAI (2024) and previous works Nori et al. (2023). Despite employing techniques like few-shot prompting with state-of-the-art LLMs, we still observe mistakes, indicating the complexity of this problem. To address this, we conducted manual checks to minimize errors and share the dataset with the research community for further improvement. 9 Ethical Concerns Our work presents a scalable and interpretable framework for fact-checking textual claims. To promote the exploration of this important problem space, we fine-tune and share open-source LLMs that are well-aligned to the framework’s objective function. While the models we provide perform better than the publicly available base models for our specific task, they still share similar weaknesses such as erroneous reasoning. To address this as best as we can, we have incorporated and described ways to mitigate these issues to the extent possible. We believe the benefits of this work outweigh any potential risks. References - Achiam et al. (2023) Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774. - Aly et al. (2021) Rami Aly, Zhijiang Guo, Michael Schlichtkrull, James Thorne, Andreas Vlachos, Christos Christodoulopoulos, Oana Cocarascu, and Arpit Mittal. 2021. Feverous: Fact extraction and verification over unstructured and structured information. arXiv preprint arXiv:2106.05707. - Bohnet et al. (2022) Bernd Bohnet, Vinh Q Tran, Pat Verga, Roee Aharoni, Daniel Andor, Livio Baldini Soares, Massimiliano Ciaramita, Jacob Eisenstein, Kuzman Ganchev, Jonathan Herzig, et al. 2022. Attributed question answering: Evaluation and modeling for attributed large language models. arXiv preprint arXiv:2212.08037. - Chen et al. (2019) Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. 2019. Tabfact: A large-scale dataset for table-based fact verification. arXiv preprint arXiv:1909.02164. - Chowdhery et al. (2023) Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1–113. - Conneau et al. (2017) Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic Barrault, and Antoine Bordes. 2017. Supervised learning of universal sentence representations from natural language inference data. arXiv preprint arXiv:1705.02364. - Dettmers et al. (2024) Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2024. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36. - Dong and Smith (2021) Rui Dong and David A Smith. 2021. Structural encoding and pre-training matter: Adapting bert for table-based fact verification. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2366–2375. - Gao et al. (2023) Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, et al. 2023. Rarr: Researching and revising what language models say, using language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16477–16508. - Gerber (2023) Emanuel Gerber. 2023. spacy module for linking text to wikidata items. https://github.com/egerber/spaCy-entity-linker. Accessed: 2024-02-26. - Guan et al. (2023) Jian Guan, Jesse Dodge, David Wadden, Minlie Huang, and Hao Peng. 2023. Language models hallucinate, but may excel at fact verification. arXiv preprint arXiv:2310.14564. - Gutiérrez and Patricio (2023) Torres Gutiérrez and Cristóbal Patricio. 2023. Sistema visual para explorar subgrafos temáticos en wikidata. - Hu et al. (2021) Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685. - Hwang et al. (2020) Jena D. Hwang, Chandra Bhagavatula, Ronan Le Bras, Jeff Da, Keisuke Sakaguchi, Antoine Bosselut, and Yejin Choi. 2020. Comet-atomic 2020: On symbolic and neural commonsense knowledge graphs. In AAAI Conference on Artificial Intelligence. - Javaheripi et al. (2023) Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al. 2023. Phi-2: The surprising power of small language models. Microsoft Research Blog. - Jiang et al. (2023a) Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023a. Mistral 7b. arXiv preprint arXiv:2310.06825. - Jiang et al. (2023b) Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023b. Active retrieval augmented generation. arXiv preprint arXiv:2305.06983. - Kim et al. (2023) Dahyun Kim, Chanjun Park, Sanghoon Kim, Wonsung Lee, Wonho Song, Yunsu Kim, Hyeonwoo Kim, Yungi Kim, Hyeonju Lee, Jihoo Kim, et al. 2023. Solar 10.7 b: Scaling large language models with simple yet effective depth up-scaling. arXiv preprint arXiv:2312.15166. - Kojima et al. (2022) Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199–22213. - Lee et al. (2017) Kenton Lee, Luheng He, Mike Lewis, and Luke Zettlemoyer. 2017. End-to-end neural coreference resolution. arXiv preprint arXiv:1707.07045. - Lee et al. (2020) Nayeon Lee, Belinda Z Li, Sinong Wang, Wen-tau Yih, Hao Ma, and Madian Khabsa. 2020. Language models as fact checkers? arXiv preprint arXiv:2006.04102. - Lin (2004) Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics. - Mallen et al. (2022) Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Hannaneh Hajishirzi, and Daniel Khashabi. 2022. When not to trust language models: Investigating effectiveness and limitations of parametric and non-parametric memories. arXiv preprint arXiv:2212.10511. - Min et al. (2023) Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. Factscore: Fine-grained atomic evaluation of factual precision in long form text generation. Preprint, arXiv:2305.14251. - Monteith et al. (2024) Scott Monteith, Tasha Glenn, John R Geddes, Peter C Whybrow, Eric Achtyes, and Michael Bauer. 2024. Artificial intelligence and increasing misinformation. The British Journal of Psychiatry, 224(2):33–35. - Nori et al. (2023) Harsha Nori, Yin Tat Lee, Sheng Zhang, Dean Carignan, Richard Edgar, Nicolo Fusi, Nicholas King, Jonathan Larson, Yuanzhi Li, Weishung Liu, et al. 2023. Can generalist foundation models outcompete special-purpose tuning? case study in medicine. arXiv preprint arXiv:2311.16452. - OpenAI (2024) OpenAI. 2024. Best practices for prompt engineering with the openai api. https://help.openai.com/en/articles/6654000-best-practices-for-prompt-engineering-with-the-openai-api. Accessed:2024-01-11. - Rashkin et al. (2023) Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter. 2023. Measuring attribution in natural language generation models. Computational Linguistics, pages 1–64. - Rechkemmer and Yin (2022) Amy Rechkemmer and Ming Yin. 2022. When confidence meets accuracy: Exploring the effects of multiple performance indicators on trust in machine learning models. In Proceedings of the 2022 chi conference on human factors in computing systems, pages 1–14. - Russell et al. (2015) Ginny Russell, Stephan Collishaw, Jean Golding, Susan E Kelly, and Tamsin Ford. 2015. Changes in diagnosis rates and behavioural traits of autism spectrum disorder over time. BJPsych open, 1(2):110–115. - Schick and Schütze (2020) Timo Schick and Hinrich Schütze. 2020. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676. - Sciavolino et al. (2021) Christopher Sciavolino, Zexuan Zhong, Jinhyuk Lee, and Danqi Chen. 2021. Simple entity-centric questions challenge dense retrievers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6138–6148, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. - Shin (2021) Donghee Shin. 2021. The effects of explainability and causability on perception, trust, and acceptance: Implications for explainable ai. International Journal of Human-Computer Studies, 146:102551. - Team et al. (2024) Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. 2024. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295. - Thorne et al. (2018) James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. Fever: a large-scale dataset for fact extraction and verification. arXiv preprint arXiv:1803.05355. - Thorne et al. (2019) James Thorne, Andreas Vlachos, Oana Cocarascu, Christos Christodoulopoulos, and Arpit Mittal. 2019. The FEVER2.0 shared task. In Proceedings of the Second Workshop on Fact Extraction and VERification (FEVER), pages 1–6, Hong Kong, China. Association for Computational Linguistics. - Touvron et al. (2023) Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288. - Tunstall et al. (2023) Lewis Tunstall, Edward Beeching, Nathan Lambert, Nazneen Rajani, Kashif Rasul, Younes Belkada, Shengyi Huang, Leandro von Werra, Clémentine Fourrier, Nathan Habib, et al. 2023. Zephyr: Direct distillation of lm alignment. arXiv preprint arXiv:2310.16944. - Vrandečić and Krötzsch (2014) Denny Vrandečić and Markus Krötzsch. 2014. Wikidata: a free collaborative knowledgebase. Communications of the ACM, 57(10):78–85. - Wang et al. (2021a) Nancy XR Wang, Diwakar Mahajan, Marina Danilevsky, and Sara Rosenthal. 2021a. Semeval-2021 task 9: Fact verification and evidence finding for tabular data in scientific documents (sem-tab-facts). arXiv preprint arXiv:2105.13995. - Wang et al. (2021b) Sinong Wang, Han Fang, Madian Khabsa, Hanzi Mao, and Hao Ma. 2021b. Entailment as few-shot learner. arXiv preprint arXiv:2104.14690. - Weitz et al. (2019) Katharina Weitz, Dominik Schiller, Ruben Schlagowski, Tobias Huber, and Elisabeth André. 2019. " do you trust me?" increasing user-trust by integrating virtual agents in explainable ai interaction design. In Proceedings of the 19th ACM International Conference on Intelligent Virtual Agents, pages 7–9. - Yang et al. (2015) Yi Yang, Wen-tau Yih, and Christopher Meek. 2015. WikiQA: A challenge dataset for open-domain question answering. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2013–2018, Lisbon, Portugal. Association for Computational Linguistics. - Yang et al. (2018) Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600. - Yao et al. (2022) Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629. - Yue et al. (2023) Xiang Yue, Boshi Wang, Kai Zhang, Ziru Chen, Yu Su, and Huan Sun. 2023. Automatic evaluation of attribution by large language models. arXiv preprint arXiv:2305.06311. - Zhou et al. (2023) Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. 2023. Lima: Less is more for alignment. arXiv preprint arXiv:2305.11206. Appendix A Appendix A.1 Training Details In this section, we present the training parameters used for fine-tuning each model, along with their corresponding loss plots. All models converged after two epochs, achieving ROUGE-L Lin (2004) scores greater than 0.658, with the best model reaching 0.719. <details> <summary>x4.png Details</summary>  ### Visual Description ## Line Chart: Llama3-8B-Chat Loss Over Epochs ### Overview The chart visualizes the training loss of the Llama3-8B-Chat model across 2.00 epochs. Two lines are plotted: "Original" (dark purple) and "Smoothed" (light pink), showing loss values on the y-axis against epoch progression on the x-axis. Both lines exhibit a general downward trend, with the Smoothed line consistently maintaining lower loss values than the Original line after the initial drop. ### Components/Axes - **Title**: "Llama3-8B-Chat" (top-center, bold black text). - **Legend**: Located in the top-right corner, with: - "Original" (dark purple line). - "Smoothed" (light pink line). - **X-axis (Epoch)**: - Label: "Epoch" (bottom-center, black text). - Scale: 0.00 to 2.00 in increments of 0.25. - **Y-axis (Loss)**: - Label: "Loss" (left-center, black text). - Scale: 0.00 to 2.00 in increments of 0.25. ### Detailed Analysis 1. **Original Line**: - Starts at **1.00 loss** at epoch 0.00. - Drops sharply to **~0.30 loss** by epoch 0.50. - Exhibits minor fluctuations (~0.30–0.40) between epochs 0.50–2.00. - Final loss value: **~0.30** at epoch 2.00. 2. **Smoothed Line**: - Starts at **~0.95 loss** at epoch 0.00. - Declines steadily to **~0.35 loss** by epoch 2.00. - Maintains a smoother trajectory with no sharp drops or spikes. - Final loss value: **~0.35** at epoch 2.00. ### Key Observations - The Smoothed line consistently outperforms the Original line in terms of lower loss after the initial epoch. - Both lines converge toward similar loss values (~0.30–0.35) by epoch 2.00, suggesting stabilization. - The Original line shows higher variance in early epochs (e.g., sharp drop from 1.00 to 0.30 between epochs 0.00–0.50), while the Smoothed line exhibits a more gradual decline. ### Interpretation The chart demonstrates the effectiveness of smoothing techniques in reducing loss volatility during model training. The Original line’s initial sharp drop may reflect rapid learning, but subsequent fluctuations suggest instability or overfitting. The Smoothed line’s consistent performance indicates that smoothing mitigates noise, leading to more reliable training. The convergence of both lines by epoch 2.00 implies that the model reaches a stable state, with smoothing providing a more robust path to convergence. This aligns with common practices in machine learning, where smoothing (e.g., gradient clipping or moving averages) is used to improve training stability. </details> Figure 4: Fine-tuning loss plots for Llama3-8B-Chat. | Parameter Base Model ROUGE-L | Value meta-llama/Meta-Llama-3-8B-Instruct 0.679 | | --- | --- | | ROUGE-1 | 0.705 | | Fine-Tuning Type | LoRA | | LoRA Alpha | 16 | | LoRA Rank | 8 | | Cutoff Length | 4096 | | Gradient Accumulation Steps | 8 | | Learning Rate | 5.0e-05 | | LR Scheduler Type | Cosine | | Number of Training Epochs | 2.0 | | Optimizer | AdamW | | Quantization Bit | 4 | Table 5: Fine-tuning Parameters for Llama3-8B-Chat <details> <summary>x5.png Details</summary>  ### Visual Description ## Line Chart: Mistral-7B-v0.3-Chat ### Overview The chart visualizes the training loss of the Mistral-7B-v0.3-Chat model across 2.00 epochs. Two lines are plotted: "Original" (dark purple) and "Smoothed" (light purple), showing loss reduction over time. Both lines converge toward lower loss values, indicating model improvement during training. ### Components/Axes - **Title**: "Mistral-7B-v0.3-Chat" (top-center, bold black text). - **X-axis**: "Epoch" (horizontal, linear scale from 0.00 to 2.00). - **Y-axis**: "Loss" (vertical, linear scale from 0.00 to 2.00). - **Legend**: Top-right corner, with labels: - "Original" (dark purple line). - "Smoothed" (light purple line). ### Detailed Analysis 1. **Original Line**: - Starts at **~0.75 loss** at epoch 0.00. - Sharp decline to **~0.25 loss** by epoch 0.5. - Stabilizes between **0.25–0.30 loss** from epoch 0.5 to 2.00. - Minor fluctuations (e.g., slight dip to 0.20 at epoch 1.75). 2. **Smoothed Line**: - Begins at **~0.70 loss** at epoch 0.00. - Gradual decline to **~0.25 loss** by epoch 1.5. - Remains stable between **0.25–0.30 loss** from epoch 1.5 to 2.00. - Smoother trajectory with no sharp drops. ### Key Observations - Both lines converge to similar loss values (~0.25–0.30) by epoch 2.00. - The "Original" line exhibits higher volatility, while the "Smoothed" line shows a more consistent decline. - Loss reduction is most pronounced in the first 0.5 epochs for both lines. ### Interpretation The chart demonstrates that the Mistral-7B-v0.3-Chat model's training loss decreases significantly during the initial epochs, with both "Original" and "Smoothed" approaches achieving comparable performance by epoch 2.00. The "Smoothed" line likely represents an averaged or filtered metric, reducing noise in the loss calculation. The convergence suggests that both training strategies are effective, but the "Smoothed" line may provide a more stable estimate of the model's learning progress. The sharp early decline indicates rapid initial learning, while the plateau suggests diminishing returns as the model approaches optimal performance. </details> Figure 5: Fine-tuning loss plots for Mistral-7B-v0.3-Chat. | Parameter Base Model ROUGE-L | Value mistralai/Mistral-7B-Instruct-v0.3 0.694 | | --- | --- | | ROUGE-1 | 0.719 | | Fine-Tuning Type | LoRA | | LoRA Alpha | 16 | | LoRA Rank | 8 | | Cutoff Length | 4096 | | Gradient Accumulation Steps | 8 | | Learning Rate | 5.0e-05 | | LR Scheduler Type | Cosine | | Number of Training Epochs | 2.0 | | Optimizer | AdamW | | Quantization Bit | 4 | Table 6: Fine-tuning Parameters for Mistral-7B-v0.3-Chat <details> <summary>x6.png Details</summary>  ### Visual Description ## Line Chart: Phi-3-mini-4k-Chat Loss Over Epochs ### Overview The chart visualizes the loss reduction of two model configurations ("Original" and "Smoothed") over 2.00 training epochs. Both lines show a downward trend, with the "Smoothed" line exhibiting less fluctuation than the "Original" line. ### Components/Axes - **X-axis (Epoch)**: Ranges from 0.00 to 2.00 in increments of 0.25. - **Y-axis (Loss)**: Ranges from 0.00 to 2.00 in increments of 0.25. - **Legend**: Located in the top-right corner, with: - **Orange line**: "Original" model - **Light orange line**: "Smoothed" model ### Detailed Analysis 1. **Original Line (Orange)**: - Starts at ~0.8 loss at epoch 0.00. - Decreases to ~0.3 loss by epoch 2.00. - Exhibits minor fluctuations (e.g., slight dips at epochs 0.50 and 1.25). - Average slope: ~-0.25 loss per epoch. 2. **Smoothed Line (Light Orange)**: - Starts at ~0.85 loss at epoch 0.00. - Decreases to ~0.35 loss by epoch 2.00. - Maintains a smoother trajectory with no sharp fluctuations. - Average slope: ~-0.275 loss per epoch. 3. **Convergence**: - Both lines converge near epoch 2.00, with the "Smoothed" line ending slightly higher (~0.35 vs. ~0.3). ### Key Observations - The "Smoothed" line demonstrates reduced volatility, suggesting improved stability in loss tracking. - Both models achieve ~65-70% loss reduction over 2 epochs. - Early epochs (0.00–0.50) show the steepest decline for both lines. ### Interpretation The "Smoothed" line’s reduced fluctuations imply that smoothing techniques effectively filter noise, providing a clearer view of the underlying trend. The convergence at epoch 2.00 suggests both configurations reach similar performance levels by the end of training. The "Original" line’s minor fluctuations may indicate transient instability during training, which smoothing mitigates. This aligns with common practices in machine learning where smoothing aids in diagnosing model behavior without overfitting to noise. </details> Figure 6: Fine-tuning loss plots for Phi-3-mini-4k-Chat. | Parameter Base Model ROUGE-L | Value microsoft/Phi-3-mini-4k-instruct 0.658 | | --- | --- | | ROUGE-1 | 0.685 | | Fine-Tuning Type | LoRA | | LoRA Alpha | 16 | | LoRA Rank | 8 | | Cutoff Length | 4096 | | Gradient Accumulation Steps | 8 | | Learning Rate | 5.0e-05 | | LR Scheduler Type | Cosine | | Number of Training Epochs | 2.0 | | Optimizer | AdamW | | Quantization Bit | 4 | Table 7: Fine-tuning Parameters for Phi-3-mini-4k-Chat <details> <summary>x7.png Details</summary>  ### Visual Description ## Line Chart: Solar-10.7B-Chat Loss Over Epochs ### Overview The chart visualizes the loss reduction of a model named "Solar-10.7B-Chat" across 2.00 epochs. Two lines are plotted: "Original" (black) and "Smoothed" (gray), showing distinct trends in loss values over time. ### Components/Axes - **X-axis (Epoch)**: Ranges from 0.00 to 2.00 in increments of 0.25. - **Y-axis (Loss)**: Ranges from 0.00 to 2.00 in increments of 0.25. - **Legend**: Located in the top-right corner, with "Original" (black line) and "Smoothed" (gray line) labels. - **Title**: "Solar-10.7B-Chat" is centered at the top of the chart. ### Detailed Analysis 1. **Original Line (Black)**: - Starts at **~0.75 loss** at epoch 0.00. - Drops sharply to **~0.25 loss** by epoch 0.5. - Fluctuates slightly between **0.20–0.30 loss** from epoch 0.5 to 2.00. - Ends at **~0.25 loss** at epoch 2.00. 2. **Smoothed Line (Gray)**: - Begins at **~0.75 loss** at epoch 0.00. - Decreases gradually to **~0.25 loss** by epoch 2.00. - Maintains a steady decline with minimal fluctuations. - Ends at **~0.25 loss** at epoch 2.00. ### Key Observations - The "Original" line exhibits a **steep initial decline** followed by stabilization, while the "Smoothed" line shows a **consistent, gradual decrease**. - Both lines converge near **0.25 loss** by epoch 2.00, suggesting alignment in long-term performance. - The "Smoothed" line’s reduced volatility implies it filters short-term noise in the loss metric. ### Interpretation The data demonstrates that the "Original" loss metric experiences rapid improvement early in training but plateaus, whereas the "Smoothed" metric provides a more stable representation of progress. This suggests that smoothing techniques effectively mitigate epoch-to-epoch variability, offering a clearer view of sustained model performance. The convergence at epoch 2.00 indicates that both metrics ultimately reflect similar loss values, though the "Smoothed" line’s gradual decline may better represent long-term optimization trends. </details> Figure 7: Fine-tuning loss plots for SOLAR-10.7B-Chat. | Parameter Base Model ROUGE-L | Value upstage/SOLAR-10.7B-Instruct-v1.0 0.689 | | --- | --- | | ROUGE-1 | 0.714 | | Fine-Tuning Type | LoRA | | LoRA Alpha | 16 | | LoRA Rank | 8 | | Cutoff Length | 4096 | | Gradient Accumulation Steps | 8 | | Learning Rate | 5.0e-05 | | LR Scheduler Type | Cosine | | Number of Training Epochs | 2.0 | | Optimizer | AdamW | | Quantization Bit | 4 | Table 8: Fine-tuning Parameters for SOLAR-10.7B-Chat <details> <summary>x8.png Details</summary>  ### Visual Description ## Line Chart: Vicuna-7B-v1.5-Chat Training Loss ### Overview The chart visualizes the training loss of the Vicuna-7B-v1.5-Chat model across 2.0 epochs. Two lines are plotted: "Original" (red) and "Smoothed" (light red), showing loss reduction over time. Both lines converge toward lower loss values, indicating model stabilization. ### Components/Axes - **X-axis (Epoch)**: Labeled "Epoch," with markers at 0.00, 0.25, 0.50, 0.75, 1.00, 1.25, 1.50, 1.75, and 2.00. - **Y-axis (Loss)**: Labeled "Loss," with markers at 0.00, 0.25, 0.50, 0.75, 1.00, 1.25, 1.50, 1.75, and 2.00. - **Legend**: Located in the top-right corner, associating: - Red line with "Original" - Light red line with "Smoothed" ### Detailed Analysis 1. **Original Line (Red)**: - Starts at **~1.0 loss** at epoch 0.00. - Declines sharply to **~0.4 loss** by epoch 0.50. - Plateaus between **~0.3–0.4 loss** from epoch 0.75 to 2.00, with minor fluctuations (e.g., ~0.32 at epoch 1.75). 2. **Smoothed Line (Light Red)**: - Begins at **~0.95 loss** at epoch 0.00. - Follows a smoother decline to **~0.35 loss** by epoch 2.00. - Exhibits less variance than the Original line, with a consistent downward trend. ### Key Observations - Both lines show a **monotonic decrease** in loss, confirming model improvement over epochs. - The Smoothed line reduces noise in the Original line’s trajectory, highlighting a clearer trend. - Convergence near epoch 2.00 suggests diminishing returns in training efficiency. ### Interpretation The data demonstrates effective model training, with loss reduction indicating improved performance. The Smoothed line’s role in filtering noise emphasizes the importance of trend analysis in large-scale training. The convergence of both lines toward lower loss values implies that the model stabilizes after ~1.5 epochs, with minimal further gains expected beyond this point. This aligns with typical training dynamics where early epochs drive significant improvement, followed by stabilization. </details> Figure 8: Fine-tuning loss plots for Vicuna-7B-v1.5-Chat. | Parameter Base Model ROUGE-L | Value lmsys/vicuna-7b-v1.5 0.676 | | --- | --- | | ROUGE-1 | 0.700 | | Fine-Tuning Type | LoRA | | LoRA Alpha | 16 | | LoRA Rank | 8 | | Cutoff Length | 4096 | | Gradient Accumulation Steps | 8 | | Learning Rate | 5.0e-05 | | LR Scheduler Type | Cosine | | Number of Training Epochs | 2.0 | | Optimizer | AdamW | | Quantization Bit | 4 | Table 9: Fine-tuning Parameters for Vicuna-7B-v1.5-Chat <details> <summary>x9.png Details</summary>  ### Visual Description ## Line Chart: Zephyr-7B-Beta-Chat Training Loss ### Overview The chart visualizes the training loss of the Zephyr-7B-Beta-Chat model across 2.00 epochs. Two data series are compared: "Original" (green line) and "Smoothed" (light green line). Both lines show a general downward trend in loss, with the Original line exhibiting sharper initial declines and greater volatility compared to the Smoothed line. ### Components/Axes - **X-axis (Epoch)**: Ranges from 0.00 to 2.00 in increments of 0.25. - **Y-axis (Loss)**: Ranges from 0.00 to 2.00 in increments of 0.25. - **Legend**: Located in the top-right corner, with: - **Green line**: Labeled "Original" - **Light green line**: Labeled "Smoothed" ### Detailed Analysis 1. **Original Line (Green)**: - Starts at ~0.8 loss at epoch 0.00. - Drops sharply to ~0.3 by epoch 0.50. - Stabilizes between ~0.25–0.3 from epoch 0.75 to 2.00. - Exhibits minor fluctuations (e.g., slight peaks at epochs 0.75 and 1.25). 2. **Smoothed Line (Light Green)**: - Begins at ~0.85 loss at epoch 0.00. - Declines gradually to ~0.25 by epoch 2.00. - Maintains a smoother trajectory with no sharp drops or spikes. ### Key Observations - The Original line shows a **steeper initial decline** (~0.8 → 0.3 in 0.5 epochs) compared to the Smoothed line (~0.85 → 0.55 in the same interval). - Both lines converge to similar loss values (~0.25) by epoch 2.00. - The Smoothed line demonstrates **lower volatility**, suggesting post-processing or averaging reduced noise in the loss metric. ### Interpretation The chart indicates that the Zephyr-7B-Beta-Chat model's training loss decreases over time, with the Smoothed line reflecting a more stable and consistent improvement. The Original line's volatility suggests potential overfitting or sensitivity to training dynamics, while the Smoothed line's gradual decline implies effective regularization or averaging techniques. The convergence of both lines at epoch 2.00 highlights that the model stabilizes its performance regardless of smoothing, though the Smoothed line provides a clearer signal of long-term trends. </details> Figure 9: Fine-tuning loss plots for Zephyr-7B-Beta-Chat. | Parameter Base Model ROUGE-L | Value HuggingFaceH4/zephyr-7b-beta 0.686 | | --- | --- | | ROUGE-1 | 0.712 | | Fine-Tuning Type | LoRA | | LoRA Alpha | 16 | | LoRA Rank | 8 | | Cutoff Length | 4096 | | Gradient Accumulation Steps | 8 | | Learning Rate | 5.0e-05 | | LR Scheduler Type | Cosine | | Number of Training Epochs | 2.0 | | Optimizer | AdamW | | Quantization Bit | 4 | Table 10: Fine-tuning Parameters for Zephyr-7B-Beta-Chat <details> <summary>x10.png Details</summary>  ### Visual Description ## Line Chart: Gemma-7B-IT-Chat Loss Over Epochs ### Overview The chart visualizes the training loss of the Gemma-7B-IT-Chat model across 2.00 epochs. Two data series are plotted: "Original" (dark brown line) and "Smoothed" (light brown line). Both lines show a general decline in loss, with the "Smoothed" line exhibiting less volatility. ### Components/Axes - **Title**: "Gemma-7B-IT-Chat" (top-center) - **X-axis**: "Epoch" (0.00 to 2.00, increments of 0.25) - **Y-axis**: "Loss" (0.00 to 2.00, increments of 0.25) - **Legend**: Top-right corner, labels: - Dark brown: "Original" - Light brown: "Smoothed" ### Detailed Analysis 1. **Original Line**: - Starts at **~1.55 loss** at epoch 0.00. - Sharp decline to **~0.35 loss** by epoch 0.50. - Fluctuates slightly between **0.30–0.40 loss** from epoch 0.50 to 2.00. - Ends at **~0.30 loss** at epoch 2.00. 2. **Smoothed Line**: - Starts at **~1.50 loss** at epoch 0.00. - Gradual decline to **~0.30 loss** by epoch 2.00. - Minimal fluctuations, maintaining a steady downward trend. ### Key Observations - The "Original" line shows significant early volatility, dropping rapidly in the first 0.50 epochs. - The "Smoothed" line reduces noise, providing a clearer view of the overall loss trend. - Both lines converge near the end, ending at similar loss values (~0.30). ### Interpretation The data suggests that the model's loss decreases over epochs, with the smoothing technique mitigating short-term fluctuations. This indicates that while the "Original" line captures raw training dynamics, the "Smoothed" line better reflects the model's long-term performance trajectory. The convergence at epoch 2.00 implies both methods achieve comparable final performance, though smoothing stabilizes the loss measurement. </details> Figure 10: Fine-tuning loss plots for Gemma-7B-IT-Chat. | Parameter Base Model ROUGE-L | Value google/gemma-7b-it 0.678 | | --- | --- | | ROUGE-1 | 0.703 | | Fine-Tuning Type | LoRA | | LoRA Alpha | 16 | | LoRA Rank | 8 | | Cutoff Length | 4096 | | Gradient Accumulation Steps | 8 | | Learning Rate | 5.0e-05 | | LR Scheduler Type | Cosine | | Number of Training Epochs | 2.0 | | Optimizer | AdamW | | Quantization Bit | 4 | Table 11: Fine-tuning Parameters for Gemma-7B-IT-Chat <details> <summary>x11.png Details</summary>  ### Visual Description ## Line Chart: Gemma-2B-IT-Chat Loss Over Epochs ### Overview The chart visualizes the training loss of the Gemma-2B-IT-Chat model across 2.00 epochs. Two lines are plotted: "Original" (dark blue) and "Smoothed" (light blue). The y-axis represents loss values (0.00–2.00), while the x-axis represents epochs (0.00–2.00). The legend is positioned in the top-right corner. ### Components/Axes - **Title**: "Gemma-2B-IT-Chat" (top-center). - **X-axis**: Labeled "Epoch," with markers at 0.00, 0.25, 0.50, 0.75, 1.00, 1.25, 1.50, 1.75, and 2.00. - **Y-axis**: Labeled "Loss," with markers at 0.00, 0.25, 0.50, 0.75, 1.00, 1.25, 1.50, 1.75, and 2.00. - **Legend**: Top-right corner, with "Original" (dark blue) and "Smoothed" (light blue) labels. ### Detailed Analysis - **Original Line (Dark Blue)**: - Starts at ~1.6 loss at epoch 0.00. - Sharp decline to ~0.5 loss by epoch 0.50. - Fluctuates between ~0.4–0.5 loss from epoch 0.50 to 2.00. - Ends at ~0.4 loss at epoch 2.00. - **Smoothed Line (Light Blue)**: - Starts at ~1.5 loss at epoch 0.00. - Gradual decline to ~0.5 loss by epoch 1.50. - Stabilizes around ~0.4–0.5 loss from epoch 1.50 to 2.00. - Ends at ~0.4 loss at epoch 2.00. ### Key Observations 1. **Initial Sharp Decline**: The Original line experiences a rapid loss reduction in the first 0.5 epochs (~1.6 → ~0.5). 2. **Smoothing Effect**: The Smoothed line exhibits a more gradual decline, avoiding abrupt changes. 3. **Convergence**: Both lines converge to similar loss values (~0.4–0.5) by epoch 1.50 and remain stable thereafter. 4. **Fluctuations**: The Original line shows minor oscillations (~0.4–0.5) after epoch 0.50, while the Smoothed line remains smoother. ### Interpretation The chart demonstrates the training dynamics of the Gemma-2B-IT-Chat model. The Original line’s sharp initial drop suggests rapid early learning, while the Smoothed line’s gradual decline indicates a more stable optimization process. The convergence of both lines by epoch 1.50 implies that both approaches achieve comparable performance by the end of training. The Smoothed line’s reduced volatility may reflect techniques like moving averages or gradient clipping to mitigate noise in the loss signal. The final loss values (~0.4) suggest moderate model performance, with potential for further optimization. </details> Figure 11: Fine-tuning loss plots for Gemma-2B-IT-Chat. | Parameter Base Model ROUGE-L | Value google/gemma-2b-it 0.667 | | --- | --- | | ROUGE-1 | 0.692 | | Fine-Tuning Type | LoRA | | LoRA Alpha | 16 | | LoRA Rank | 8 | | Cutoff Length | 4096 | | Gradient Accumulation Steps | 8 | | Learning Rate | 5.0e-05 | | LR Scheduler Type | Cosine | | Number of Training Epochs | 2.0 | | Optimizer | AdamW | | Quantization Bit | 4 | Table 12: Fine-tuning Parameters for Gemma-2B-IT-Chat **Text Span Attribution Verification** **Objective:** Predict whether the text span is "Attributable", "Contradictory", or "Extrapolatory" based on the information provided in the triplets. **Instructions:** 1. **Read the Full Text:** - Understand the context and content of the full text string. 2. **Examine the Text Span:** - Determine the claims made within the text span. 3. **Analyze the Triplets:** - Evaluate if the triplets support, refute, or neither support nor refute the claims in the text span. 4. **Make Your Prediction:** - Classify the text span as "Attributable", "Contradictory", or "Extrapolatory" based on your analysis of the triplets. 5. **Provide Rationale:** - Clearly explain your reasoning for the classification. **Classification Criteria:** - **"Attributable"**: The text span is sufficiently supported by the triplet(s). All claims in the text span are directly present in the triplet information. - **"Contradictory"**: The text span is conclusively refuted by the triplet(s). All claims in the text span are directly contradicted by the triplet information. - **"Extrapolatory"**: The triplet(s) can neither support nor refute the text span. The information provided is either irrelevant, indirect, or related but not sufficient to support or refute the text span. **Example:** **Full Text:** "Albert Einstein is widely recognized as the father of modern physics. He was awarded the Nobel Prize in Physics for his services to Theoretical Physics." **Text Span:** "He was awarded the Nobel Prize in Physics." **Triplets:** [("Albert Einstein", "award received", "Nobel Prize in Physics")] **Sample Evaluation:** - **Prediction:** "Attributable" - **Rationale:** "The triplet directly supports the claim that Albert Einstein received the Nobel Prize in Physics." **Example:** **Full Text:** "Isaac Newton discovered the element radium." **Text Span:** "Isaac Newton discovered radium." **Triplets:** [("Marie Curie", "discovered", "radium")] **Sample Evaluation:** - **Prediction:** "Contradictory" - **Rationale:** "The triplet states that Marie Curie discovered radium, contradicting the claim that Isaac Newton discovered it." **Example:** **Full Text:** "The Eiffel Tower is a wrought-iron lattice tower that was opened in 1889." **Text Span:** "The Eiffel Tower is a wrought-iron lattice tower that was opened in 1889." **Triplets:** [("Eiffel Tower", "located in", "Paris")] **Sample Evaluation:** - **Prediction:** "Extrapolatory" - **Rationale:** "The triplet states that the Eiffel Tower is located in Paris, which is related but not sufficient to confirm or refute that it was opened in 1889." **Verification Checklist:** - [ ] The prediction accurately reflects the relationship between the text span and the triplets. - [ ] The rationale clearly explains the classification based on the triplets. - [ ] The explanation is free from irrelevant information. **Response Format:** Provide your evaluation in the following JSON format: - "prediction": "Attributable", "Contradictory", or "Extrapolatory" - "rationale": "Your comments here" **Inputs to Evaluate** **Full text:** "{full_text}" **Text span:** "{text_span}" **Triplets:** {triplets} Figure 12: Prompt for the text span attribution verification task, guiding the model to classify text spans as "Attributable," "Contradictory," or "Extrapolatory" based on the provided triplets. The prompt design incorporates concepts such as few-shot learning, chain-of-thought reasoning Kojima et al. (2022), and tailored prompt engineering OpenAI (2024); Nori et al. (2023)