Datasets:

alecocc
/

mmscifact-dev

qa_pair_id stringlengths 29 89	scenario stringclasses 4 values	level stringclasses 3 values	question_subtype stringclasses 9 values	answer_model stringlengths 12 50	question_text stringlengths 123 589	question_meta dict	annotations listlengths 5 25	source_images listlengths 0 4	answer_raw stringlengths 789 4.45k	question_model stringclasses 6 values	paper_id stringclasses 6 values	paper_pdf_path stringclasses 6 values
2024_acl-long_533_cross_g2_p5_7_24_results_interpretation	S2	cross_page	results_interpretation	gemma-4-31b-it	What pattern do Table 2 and Table 11 together show about the generated-context setup relative to the retrieved-context baselines on the same benchmark, and which named rows stand out as the main exceptions or near-ties?	{ "input_types": [ "Table 2", "Table 11", "Section 5.2" ], "question_type": "results_interpretation", "visual_subtype": "quantitative", "error_subtype": null, "scope": null }	[ { "sentence": "Together, Table 2 and Table 11 demonstrate a consistent pattern where the generated-context setup, implemented through the MedGENIE framework, generally outperforms retrieved-context baselines across the evaluated benchmarks.", "label": "contradiction", "role": "Interpretation", "rati...	[ "pdfs_jpg/2024_acl-long_533/page_005.jpg", "pdfs_jpg/2024_acl-long_533/page_007.jpg", "pdfs_jpg/2024_acl-long_533/page_024.jpg" ]	Together, Table 2 and Table 11 demonstrate a consistent pattern where the generated-context setup, implemented through the MedGENIE framework, generally outperforms retrieved-context baselines across the evaluated benchmarks. In Table 11, which focuses specifically on MedQA, every MedGENIE model—ranging from MedGENIE-P...	gpt-5.4-mini	2024_acl-long_533	pdfs/2024_acl-long_533.pdf
2024_acl-long_533_full_cross_adversarial	S1	full_paper	results_interpretation	gemma-4-31b-it	I've been examining the subject-specific performance charts in Figure 6 to see how the framework handles different medical specialties. If I'm reading this right, given the abstract's claim that MEDGENIE sets a new state-of-the-art in each testbed, it looks like the FID-Flan-T5 model actually maintains a performance le...	{ "input_types": [ "Abstract", "Figure 6" ], "question_type": "cross_adversarial", "visual_subtype": "comparison", "error_subtype": "overgeneralized_scope", "scope": null }	[ { "sentence": "The interpretation that the MedGENIE-FID-Flan-T5 model maintains a performance lead over the Meditron baseline across every single subject area in MedMCQA is incorrect.", "label": "not_a_claim", "role": "no_rad", "rationale": "", "depends_on": [], "errors": [], "2d_box": [...	[]	The interpretation that the MedGENIE-FID-Flan-T5 model maintains a performance lead over the Meditron baseline across every single subject area in MedMCQA is incorrect. While the model demonstrates superior performance in many categories, the data in Figure 6 shows that it does not outperform the Meditron baseline in e...	gemini-3-flash-preview	2024_acl-long_533	pdfs/2024_acl-long_533.pdf
2024_acl-long_533_page_001_trend	S4	single_page	trend	gemini-3-flash-preview	In Figure 1, how does accuracy change over publication date for the model series—does it climb steadily, flatten out, dip at any point, or jump around—and where are the main rises or drops most noticeable?	{ "input_types": [ "Figure 1" ], "question_type": "trend", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "According to Figure 1, the accuracy of the model series on the MedQA-USMLE 4-opt benchmark follows a consistent upward trajectory from January 2021 to February 2024.", "label": "contradiction", "role": "Interpretation", "rationale": "", "depends_on": [ "S3", "S5", ...	[ "pdfs_jpg/2024_acl-long_533/page_001.jpg" ]	According to Figure 1, the accuracy of the model series on the MedQA-USMLE 4-opt benchmark follows a consistent upward trajectory from January 2021 to February 2024. The data illustrates a steady climb in performance over this three-year period, with the frontier line connecting the top-performing models showing no ins...	gpt-5.4-mini	2024_acl-long_533	pdfs/2024_acl-long_533.pdf
2024_acl-long_533_page_007_comparison	S4	single_page	comparison	gemma-4-31b-it	How do LLaMA-2-chat and Zephyr-beta compare across the three grounding conditions (None, Generated, and Retrieved) in Table 2? Describe where the performance gap between the two models is widest, where it narrows, and how the effectiveness of the grounding methods varies for each model across the three benchmarks.	{ "input_types": [ "Table 2", "Section 5.2" ], "question_type": "comparison", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "Across the three grounding conditions, Zephyr-β consistently outperforms LLaMA-2-chat.", "label": "supported", "role": "Interpretation", "rationale": "", "depends_on": [ "S2", "S3", "S4" ], "errors": [], "2d_box": [], "propagation_type": "none" ...	[ "pdfs_jpg/2024_acl-long_533/page_007.jpg" ]	Across the three grounding conditions, Zephyr-β consistently outperforms LLaMA-2-chat. In the no grounding condition, Zephyr-β achieves an average score of 51.1 compared to 40.4 for LLaMA-2-chat. With retrieved grounding, Zephyr-β maintains its lead with an average score of 54.8, while LLaMA-2-chat scores 42.1. Under g...	gemini-3-flash-preview	2024_acl-long_533	pdfs/2024_acl-long_533.pdf
2024_acl-long_533_page_008_quantitative	S4	single_page	quantitative	gpt-5.4-mini	Looking at Table 3, how does adding different amounts of artificial contexts (G) to the baseline knowledge base (KB) across the different models (llama2, mistral, zephyr) and datasets (MedQA, MedMCQA, MMLU) generally affect the accuracy scores?	{ "input_types": [ "Table 3" ], "question_type": "quantitative", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "Table 3 shows a broadly positive effect from adding artificial contexts G* to the baseline knowledge base KB* across all three models and across the three datasets, although the size of the gain depends on both the model and the dataset.", "label": "contradiction", "role": "Interpretation...	[ "pdfs_jpg/2024_acl-long_533/page_008.jpg" ]	Table 3 shows a broadly positive effect from adding artificial contexts G* to the baseline knowledge base KB* across all three models and across the three datasets, although the size of the gain depends on both the model and the dataset. In every setting, moving from KB* alone to KB* plus artificial contexts improves a...	gemma-4-31b-it	2024_acl-long_533	pdfs/2024_acl-long_533.pdf
2024_acl-long_533_page_022_adversarial	S5	single_page	adversarial	gemma-4-31b-it	If I'm interpreting the charts in Figure 6 correctly, the model that achieves the highest accuracy in every subject for MedMCQA maintains that top position in the MMLU benchmarks as well, specifically outperforming all other models in the Medical Genetics category.	{ "input_types": [ "Figure 6 (MedMCQA panels)", "Figure 6 (MMLU panel)" ], "question_type": "adversarial", "visual_subtype": "comparison", "error_subtype": "reversed_relationship", "scope": null }	[ { "sentence": "The interpretation of the charts in Figure 6 is incorrect.", "label": "contradiction", "role": "Interpretation", "rationale": "", "depends_on": [ "S2" ], "errors": [ { "span": "", "tags": [ "C4" ], "correction": "" ...	[ "pdfs_jpg/2024_acl-long_533/page_022.jpg" ]	The interpretation of the charts in Figure 6 is incorrect. No single model achieves the highest accuracy across every subject for the MedMCQA benchmarks. While MedGENIE-FID-Flan-T5 holds the top position in many categories, it does not do so in Ophthalmology, where MedGENIE-Zephyr-β achieves the highest accuracy at 70%...	gemini-3-flash-preview	2024_acl-long_533	pdfs/2024_acl-long_533.pdf
Beyond_Outcome_Verification_cross_g3_p5_13_14_implication	S2	cross_page	implication	google/gemma-4-31b-it	How do the zero-valued categories for the RoBBR datasets in Table 6, when mapped to their definitions in Figure 5, limit the benchmark evaluation of the model's reasoning capabilities across all risk-of-bias domains?	{ "input_types": [ "Table 1", "Table 6", "Figure 5", "Figure 6" ], "question_type": "implication", "visual_subtype": "comparison", "error_subtype": null, "scope": null }	[ { "sentence": "The zero-valued categories for the RoBBR datasets in Table 6 limit the benchmark evaluation by creating gaps where the model's reasoning capabilities cannot be assessed for specific risk-of-bias domains.", "label": "contradiction", "role": "Interpretation", "rationale": "False because...	[ "pdfs_jpg/Beyond_Outcome_Verification/page_005.jpg", "pdfs_jpg/Beyond_Outcome_Verification/page_013.jpg", "pdfs_jpg/Beyond_Outcome_Verification/page_014.jpg" ]	The zero-valued categories for the RoBBR datasets in Table 6 limit the benchmark evaluation by creating gaps where the model's reasoning capabilities cannot be assessed for specific risk-of-bias domains. According to Table 6, the RoBBR Cochrane dataset contains zero instances for categories G, H, and I, while the RoBBR...	gemini-3.5-flash	Beyond_Outcome_Verification	pdfs/Beyond_Outcome_Verification.pdf
Beyond_Outcome_Verification_full_conceptual	S1	full_paper	conceptual	qwen/qwen3.5-9b	What does Figure 2 show about how the step-by-step outputs in the model's structured reasoning template are verified and scored individually under the process rewarding scheme?	{ "input_types": [ "Figure 2", "Figure 6" ], "question_type": "conceptual", "visual_subtype": "visual_understanding", "error_subtype": null, "scope": null }	[ { "sentence": "Figure 2 provides a side-by-side comparison of \"Verifiable Outcome Rewarding\" on the left and \"Verifiable Process Rewarding\" on the right.", "label": "supported", "role": "Observation", "rationale": "", "depends_on": [], "errors": [], "2d_box": [ { "page_...	[]	Figure 2 provides a side-by-side comparison of "Verifiable Outcome Rewarding" on the left and "Verifiable Process Rewarding" on the right. The right side specifically illustrates how the VPRM framework handles step-by-step outputs. It depicts a structured reasoning trace divided into distinct steps (labeled Step 1, Ste...	gemini-3.5-flash	Beyond_Outcome_Verification	pdfs/Beyond_Outcome_Verification.pdf
Beyond_Outcome_Verification_full_methodological	S1	full_paper	methodological	qwen/qwen3.5-9b	How does the step-wise verification design in the method connect to the performance patterns in the results, and what seems to explain the improvement it produces?	{ "input_types": [ "Figure 1", "Table 4", "Table 5" ], "question_type": "methodological", "visual_subtype": "comparison", "error_subtype": null, "scope": null }	[ { "sentence": "The method’s core design involves Verifiable Process Reward Models (VPRMs), which validate every intermediate reasoning step against deterministic, domain-specific rules rather than relying solely on the final outcome.", "label": "supported", "role": "Observation", "rationale": "", ...	[]	The method’s core design involves Verifiable Process Reward Models (VPRMs), which validate every intermediate reasoning step against deterministic, domain-specific rules rather than relying solely on the final outcome. This step-wise verification design directly connects to the observed performance patterns, specifical...	gpt-5.4-mini	Beyond_Outcome_Verification	pdfs/Beyond_Outcome_Verification.pdf
Beyond_Outcome_Verification_page_002_visual_understanding	S4	single_page	visual_understanding	mistralai/mistral-small-2603	Can you walk me through what Figure 1 is showing — what are the main boxes, arrows, and labels in the top reasoning chain and the bottom decision tree, and how does each step/label lead to the final risk outcome?	{ "input_types": [ "Figure 1" ], "question_type": "visual_understanding", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "Figure 1 illustrates a two-part framework for assessing risk of bias in studies through a structured, verifiable reasoning process.", "label": "contradiction", "role": "Observation", "rationale": "This is not a \"two-part\" framework for assessing risk of bias. It's a depiction of the...	[ "pdfs_jpg/Beyond_Outcome_Verification/page_002.jpg" ]	Figure 1 illustrates a two-part framework for assessing risk of bias in studies through a structured, verifiable reasoning process. The top portion presents a sequential reasoning pipeline applied to an input study x, which involves four distinct assessment steps. Each step corresponds to a guideline-defined question a...	gpt-5.4-mini	Beyond_Outcome_Verification	pdfs/Beyond_Outcome_Verification.pdf
Beyond_Outcome_Verification_page_007_quantitative	S4	single_page	quantitative	nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free	Based on Table 2, how does the performance of the models listed under 'Our Models' change as they move from the CochraneForest benchmark to the RoBBR Cochrane and RoBBR Non-Cochrane benchmarks?	{ "input_types": [ "Table 2" ], "question_type": "quantitative", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "Based on Table 2, the performance of the models listed under \"Our Models\" consistently decreases as they transition from the CochraneForest benchmark to the RoBBR Cochrane benchmark, and continues to decline as they move to the RoBBR Non-Cochrane benchmark.", "label": "supported", "role...	[ "pdfs_jpg/Beyond_Outcome_Verification/page_007.jpg" ]	Based on Table 2, the performance of the models listed under "Our Models" consistently decreases as they transition from the CochraneForest benchmark to the RoBBR Cochrane benchmark, and continues to decline as they move to the RoBBR Non-Cochrane benchmark. Specifically, examining the Accuracy (Acc) values for each mo...	google/gemma-4-31b-it	Beyond_Outcome_Verification	pdfs/Beyond_Outcome_Verification.pdf
Beyond_Outcome_Verification_page_013_adversarial	S5	single_page	adversarial	qwen/qwen3.6-35b-a3b	If I'm reading this right, Algorithm 1 provides the core prediction logic applied to generate labels for all the risk domains (A through I) listed in Figure 5—is that correct?	{ "input_types": [ "Algorithm 1", "Figure 5" ], "question_type": "adversarial", "visual_subtype": "visual_understanding", "error_subtype": "overgeneralized_scope", "scope": null }	[ { "sentence": "Based on the provided document, it is not correct to assume that Algorithm 1 provides the core prediction logic for all risk domains listed from A through I.", "label": "not_a_claim", "role": "no_rad", "rationale": "", "depends_on": [], "errors": [], "2d_box": [], "pro...	[ "pdfs_jpg/Beyond_Outcome_Verification/page_013.jpg" ]	Based on the provided document, it is not correct to assume that Algorithm 1 provides the core prediction logic for all risk domains listed from A through I. Algorithm 1 is explicitly titled "RoB A Macro" and defines a procedure named `PREDICTLABEL-A`. This naming convention suggests the algorithm is specific to Domai...	qwen/qwen3.5-27b	Beyond_Outcome_Verification	pdfs/Beyond_Outcome_Verification.pdf
Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits_cross_g1_p4_6_8_methodological	S2	cross_page	methodological	qwen/qwen3.5-27b	What does Table 4 show about using a longer summary as input versus a predicted one when generating a shorter summary, and how does the pipeline in Figure 4b help explain that gap?	{ "input_types": [ "Figure 4b", "Table 4" ], "question_type": "methodological", "visual_subtype": "comparison", "error_subtype": null, "scope": null }	[ { "sentence": "Table 4 indicates that model performance is significantly higher when a ground truth (Gold) longer summary is provided as input compared to using a predicted (Predicted L') summary.", "label": "supported", "role": "Interpretation", "rationale": "", "depends_on": [ "S3", ...	[ "pdfs_jpg/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits/page_004.jpg", "pdfs_jpg/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits/page_006.jpg", "pdfs_jpg/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits/page_008.jpg" ]	Table 4 indicates that model performance is significantly higher when a ground truth (Gold) longer summary is provided as input compared to using a predicted (Predicted L') summary. In the progressive summarization results shown in Table 4(b), there is a notable drop in ROUGE scores when the input shifts from a Gold su...	gpt-5.4-mini	Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits	pdfs/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits.pdf
Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits_cross_g4_p5_7_9_implication	S2	cross_page	implication	nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free	Looking at Figure 1 and Table 3, what does the relationship between the target length for the shortest summaries and the actual word counts produced reveal about where the approach breaks down?	{ "input_types": [ "Figure 1", "Table 3" ], "question_type": "implication", "visual_subtype": "quantitative", "error_subtype": null, "scope": null }	[ { "sentence": "Based on Figure 1,the target length for the shortest summaries, which are labeled as \"Tiny Summaries,\" is specified as \"One sentence, ~25 words.\"", "label": "supported", "role": "Observation", "rationale": "", "depends_on": [], "errors": [], "2d_box": [ { ...	[ "pdfs_jpg/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits/page_005.jpg", "pdfs_jpg/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits/page_007.jpg", "pdfs_jpg/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits/page_009.jpg" ]	Based on Figure 1,the target length for the shortest summaries, which are labeled as "Tiny Summaries," is specified as "One sentence, ~25 words." Looking at Table 3, the task corresponding to Tiny Summaries is "D -> T" (Document to Tiny). In the "Words" column for this task, the actual word counts produced by the vari...	google/gemma-4-31b-it	Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits	pdfs/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits.pdf
Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits_page_006_quantitative	S4	single_page	quantitative	mistralai/ministral-14b-2512	Looking at Table 2, describe the variation in target words and target sentences among the datasets in the 'Short' category.	{ "input_types": [ "Table 2" ], "question_type": "quantitative", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "In the 'Short' category of Table 2, there is notable variation in the number of target words and sentences among the different datasets.", "label": "contradiction", "role": "Interpretation", "rationale": "Correct interpretation but stemming from a flawed premise. There is a hidden rea...	[ "pdfs_jpg/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits/page_006.jpg" ]	In the 'Short' category of Table 2, there is notable variation in the number of target words and sentences among the different datasets. The BigPatent dataset has a target summary consisting of 116.7 words and 3.5 sentences, which is relatively concise. In contrast, the MS^2 dataset has a significantly larger target su...	google/gemma-4-31b-it	Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits	pdfs/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits.pdf
Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits_page_009_adversarial	S5	single_page	adversarial	z-ai/glm-4.6v	If I'm reading this right, the multitask training approach always outperforms the single-task baseline in terms of R-1f1 across all the different summary length configurations mentioned in the text and Table 5. Is that right?	{ "input_types": [ "Table 5", "the paragraph above Section 5" ], "question_type": "adversarial", "visual_subtype": "comparison", "error_subtype": "overgeneralized_scope", "scope": null }	[ { "sentence": "To determine whether the multitask training approach always outperforms the single-task baseline in terms of R-1f1 across all summary length configurations, we analyze Table 5 and the accompanying text.", "label": "not_a_claim", "role": "no_rad", "rationale": "", "depends_on": [],...	[ "pdfs_jpg/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits/page_009.jpg" ]	To determine whether the multitask training approach always outperforms the single-task baseline in terms of R-1f1 across all summary length configurations, we analyze Table 5 and the accompanying text. The table compares two multitask models (one with three tasks: D → {L,S,T}; one with two tasks: L → {S,T}) against th...	google/gemma-4-31b-it	Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits	pdfs/Multi_lexsum_real_world_summaries_of_civil_rights_lawsuits.pdf
Rogue_scores_cross_g2_p3_6_19_conceptual	S2	cross_page	conceptual	google/gemma-4-26b-a4b-it	Looking at Figure 3 and Table 2, how does the error distribution of Nonstandard-Implementations explain the growth of the red bars in recent years?	{ "input_types": [ "Figure 3", "Table 2", "Section G" ], "question_type": "conceptual", "visual_subtype": "trend", "error_subtype": null, "scope": null }	[ { "sentence": "In Figure 3, the red bars represent the number of papers that cite an incorrect ROUGE package.", "label": "supported", "role": "Observation", "rationale": "Legend: red = Cites Incorrect Package (n = 755).", "depends_on": [], "errors": [], "2d_box": [ { "page_...	[ "pdfs_jpg/Rogue_scores/page_003.jpg", "pdfs_jpg/Rogue_scores/page_006.jpg", "pdfs_jpg/Rogue_scores/page_019.jpg" ]	In Figure 3, the red bars represent the number of papers that cite an incorrect ROUGE package. These bars show a significant increase in recent years, particularly from 2018 through 2022. This visual trend indicates that a growing number of research papers are utilizing ROUGE evaluation packages that result in incorrec...	google/gemma-4-31b-it	Rogue_scores	pdfs/Rogue_scores.pdf
Rogue_scores_cross_g3_p5_17_18_implication	S2	cross_page	implication	google/gemma-4-26b-a4b-it	How do Table 1 and Table C1 together show that the approach can break down as a fair way to compare models, even when the surrounding text says the score differences are often just configuration noise? Please explain what the tables reveal about the size of those hidden shifts and which kinds of settings make them swin...	{ "input_types": [ "Table 1", "Table C1" ], "question_type": "implication", "visual_subtype": "comparison", "error_subtype": null, "scope": null }	[ { "sentence": "Table C1 defines the specific experimental configurations used in the study, while Table 1 quantifies the resulting changes in ROUGE scores compared to a baseline configuration.", "label": "supported", "role": "Observation", "rationale": "The Appendix C table lists the experimental co...	[ "pdfs_jpg/Rogue_scores/page_005.jpg", "pdfs_jpg/Rogue_scores/page_017.jpg", "pdfs_jpg/Rogue_scores/page_018.jpg" ]	Table C1 defines the specific experimental configurations used in the study, while Table 1 quantifies the resulting changes in ROUGE scores compared to a baseline configuration. Together, they demonstrate that the methodology for comparing models can break down because the differences in scores caused by these configur...	gpt-5.4-mini	Rogue_scores	pdfs/Rogue_scores.pdf
Rogue_scores_full_cross_adversarial	S1	full_paper	results_interpretation	z-ai/glm-4.6v	I think the paper’s main takeaway is that ROUGE is broadly unreliable in three ways: many papers don’t report enough details to reproduce scores, evaluation choices can shift scores enough to hurt comparability, and a lot of papers rely on packages that can produce wrong results. Looking at the abstract and then Table ...	{ "input_types": [ "Abstract", "Table 2", "Figure 1" ], "question_type": "cross_adversarial", "visual_subtype": "comparison", "error_subtype": "reversed_relationship", "scope": null }	[ { "sentence": "Based strictly on the provided document, your reading is correct.", "label": "not_a_claim", "role": "no_rad", "rationale": "", "depends_on": [], "errors": [], "2d_box": [], "propagation_type": "none" }, { "sentence": "The paper’s main takeaway is that ROUGE sco...	[]	Based strictly on the provided document, your reading is correct. The paper’s main takeaway is that ROUGE scores are broadly unreliable, and the analysis of Table 2 supports the conclusion that major wrapper packages are generally more error-prone than reimplementations. The abstract explicitly states that the systema...	gpt-5.4-mini	Rogue_scores	pdfs/Rogue_scores.pdf
Rogue_scores_full_implication	S1	full_paper	implication	nvidia/nemotron-nano-12b-v2-vl:free	Where do ROUGE scores experience the most significant drops compared to the baseline under different evaluation configurations, and what does the corresponding experiment setup reveal about why those variations occur?	{ "input_types": [ "Table 1", "Appendix C" ], "question_type": "implication", "visual_subtype": "comparison", "error_subtype": null, "scope": null }	[ { "sentence": "The most significant drops in ROUGE scores compared to the baseline occur under preprocessing configurations, particularly stemming, as well as tokenization configurations like period sentence splitting and NLTK sentence splitting.", "label": "contradiction", "role": "Interpretation", ...	[]	The most significant drops in ROUGE scores compared to the baseline occur under preprocessing configurations, particularly stemming, as well as tokenization configurations like period sentence splitting and NLTK sentence splitting. The sensitivity analysis reveals that when Porter stemming is applied, ROUGE scores d...	gemini-3.5-flash	Rogue_scores	pdfs/Rogue_scores.pdf
Rogue_scores_page_001_comparison	S4	single_page	comparison	google/gemma-4-26b-a4b-it	In Figure 1, how do the percentages compare across the different study groups in panel A versus the paper practices in panel B — where are the gaps between items largest, where are they smallest, and how does the spread change overall within each panel?	{ "input_types": [ "Figure 1" ], "question_type": "comparison", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "In Panel A, which displays the reproducibility of different study groups, the data is provided in percentages: 20% for language model evaluations, 39% for psychology studies, 46% for cancer biology studies, 61% for economics studies, and 62% for social science studies.", "label": "supported",...	[ "pdfs_jpg/Rogue_scores/page_001.jpg" ]	In Panel A, which displays the reproducibility of different study groups, the data is provided in percentages: 20% for language model evaluations, 39% for psychology studies, 46% for cancer biology studies, 61% for economics studies, and 62% for social science studies. The largest gap between items in this panel is 19 ...	gpt-5.4-mini	Rogue_scores	pdfs/Rogue_scores.pdf
Rogue_scores_page_006_adversarial	S5	single_page	adversarial	nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free	I think Table 2 shows that most of the nonstandard ROUGE packages differ from ROUGE-1.5.5 across every score variant and both stemming settings, so the validation mostly finds broad, across-the-board scoring errors rather than anything localized. Is that right, or am I missing a nuance here?	{ "input_types": [ "Table 2", "Section 4.2" ], "question_type": "adversarial", "visual_subtype": "comparison", "error_subtype": "overgeneralized_scope", "scope": null }	[ { "sentence": "Based strictly on the provided document and Table 2, the statement is incorrect.", "label": "not_a_claim", "role": "no_rad", "rationale": "", "depends_on": [], "errors": [], "2d_box": [], "propagation_type": "none" }, { "sentence": "While the text notes that \"...	[ "pdfs_jpg/Rogue_scores/page_006.jpg" ]	Based strictly on the provided document and Table 2, the statement is incorrect. While the text notes that "all but one package we test has scoring errors," indicating that 18 out of 19 packages have non-zero values in at least one column, only 6 packages differ across every score variant (R1, R2, RL) and both stem...	gpt-5.4-mini	Rogue_scores	pdfs/Rogue_scores.pdf
cancers-17-02861-v3_cross_g0_p5_6_9_11_methodological	S2	cross_page	methodological	gemini-3-flash-preview	How does the clustering scheme shown in the schematic shape the survival patterns seen later, and what does the outcome figure suggest about which cluster-related pattern remains most influential after accounting for the other variables?	{ "input_types": [ "Figure 1", "Figure 2", "Section 2.5", "Section 2.7.1" ], "question_type": "methodological", "visual_subtype": "visual_understanding", "error_subtype": null, "scope": null }	[ { "sentence": "The clustering scheme presented in the document categorizes patients into four distinct transcriptomic groups—C1, C2, C3, and C4—based on their molecular signatures and underlying biological pathways.", "label": "supported", "role": "Observation", "rationale": "Correct. Evidence suppo...	[ "pdfs_jpg/cancers-17-02861-v3/page_005.jpg", "pdfs_jpg/cancers-17-02861-v3/page_006.jpg", "pdfs_jpg/cancers-17-02861-v3/page_009.jpg", "pdfs_jpg/cancers-17-02861-v3/page_011.jpg" ]	The clustering scheme presented in the document categorizes patients into four distinct transcriptomic groups—C1, C2, C3, and C4—based on their molecular signatures and underlying biological pathways. This classification directly shapes subsequent survival patterns by identifying specific molecular drivers that correla...	gpt-5.4-mini	cancers-17-02861-v3-6	pdfs/cancers-17-02861-v3-6.pdf
cancers-17-02861-v3_cross_g1_p12_13_16_results_interpretation	S2	cross_page	results_interpretation	gemma-4-31b-it	What do Figure 3, Figure 4, and the analysis described in Section 3.5 together show about the comparative performance of the various prognostic models and signatures when evaluating different survival metrics?	{ "input_types": [ "Section 3.5", "Figure 3", "Figure 4" ], "question_type": "results_interpretation", "visual_subtype": "comparison", "error_subtype": null, "scope": null }	[ { "sentence": "The provided data from Section 3.5, Figure 3, and Figure 4 collectively demonstrate that transcriptomic cluster-based classification provides superior prognostic value for overall survival (OS) compared to both traditional clinical nomograms and other molecular signatures.", "label": "support...	[ "pdfs_jpg/cancers-17-02861-v3/page_012.jpg", "pdfs_jpg/cancers-17-02861-v3/page_013.jpg", "pdfs_jpg/cancers-17-02861-v3/page_016.jpg" ]	The provided data from Section 3.5, Figure 3, and Figure 4 collectively demonstrate that transcriptomic cluster-based classification provides superior prognostic value for overall survival (OS) compared to both traditional clinical nomograms and other molecular signatures. According to the analysis in Section 3.5 and t...	gemini-3-flash-preview	cancers-17-02861-v3-6	pdfs/cancers-17-02861-v3-6.pdf

End of preview. Expand in Data Studio

MMSciFact

Multimodal scientific fact-checking benchmark: model-generated question-answer pairs over scientific PDFs, with every answer sentence human-annotated for role (Observation/Interpretation), dependency structure (depends_on), and grounding label (supported/contradiction/NEI/not_a_claim) against the source document.

40 QA pairs across 6 papers, 440 annotated sentences.
Every included QA pair has at least one contradiction/nei sentence (see src/utils/extract_final_annotations.py in the code repo).
Error taxonomy (error_tags sentence-level for C4/N4 propagation, per-span tags in errors[] for C1-C3/N1-N3) and a mechanically-derived propagation_type per sentence (false_premise / unverifiable_premise / none) — see docs/guidelines.html in the code repo for the full taxonomy definitions.

Files

mmscifact.jsonl — one row per QA pair.
pdfs.zip — one PDF per paper (pdfs/<paper_id>.pdf), referenced by paper_pdf_path on every row. Unzip next to mmscifact.jsonl before running eval scripts: unzip pdfs.zip.

Using this with the MMSciFact eval scripts

mmscifact.jsonl is a flat, single-file view of the data -- convenient for browsing, but the eval scripts in the code repo (eval_batch_api.py, run_vllm_batch.py, eval_oracle_graph.py, eval_holistic_graph.py) all expect the original one-file-per-QA-pair layout (final_annotations/<paper_id>/<qa_pair_id>.json + _source.pdf). Restore that layout with src/utils/hydrate_from_hf.py from the code repo:

hf download alecocc/mmscifact-demo --repo-type dataset --local-dir hf_download
python src/utils/hydrate_from_hf.py --dataset-dir hf_download
# writes final_annotations/<paper_id>/... at the repo root -- every eval
# script then runs completely unmodified from there.

Row schema

Each line of mmscifact.jsonl is one QA pair.

Field	Type	Meaning
`qa_pair_id`	string	Unique identifier for this QA pair
`paper_id`	string	Join key — the PDF is at `pdfs/<paper_id>.pdf` after unzipping `pdfs.zip`
`paper_pdf_path`	string	Relative path to the source PDF, `pdfs/<paper_id>.pdf`
`level`	string	How much of the paper the question requires: `single_page`, `cross_page`, or `full_paper`
`scenario`	string	Question-generation scenario code (`S2`–`S5`) — see the paper's Appendix A
`question_subtype`	string	Finer-grained question type (e.g. comparison, trend, methodological)
`question_model`	string	Model that generated the question
`answer_model`	string	Model that generated the answer being fact-checked
`question_text`	string	The question text
`question_meta`	object	Extra generation-time metadata (input types, cognitive operation, visual/error subtype, scope)
`answer_raw`	string	The full model-generated answer, before it was split into sentences
`annotations`	list of objects	Per-sentence human annotation — see below
`source_images`	list of strings	Page-number-encoding paths for which PDF page(s) the answer draws on

Each entry of annotations:

Field	Type	Meaning
`sentence`	string	The sentence text (one segment of `answer_raw`)
`label`	string	Grounding verdict: `supported`, `contradiction`, `nei`, or `not_a_claim`
`role`	string \| null	`Observation` (a direct, self-contained claim) or `Interpretation` (an inference drawn from other sentences)
`rationale`	string	Human-written explanation for the label
`depends_on`	list of strings	Sentence IDs this sentence's claim logically depends on, e.g. `["S2", "S3"]`
`error_tags`	list of strings	Sentence-level propagation code (`C4`/`N4`), if this sentence's error is inherited from a bad premise elsewhere in the answer
`propagation_type`	string	Mechanically derived from `label`+`depends_on`: `false_premise`, `unverifiable_premise`, or `none`
`errors`	list of objects	Specific erroneous spans within the sentence — see below
`2d_box`	list of objects	Bounding box(es) anchoring this sentence to the source document — see below

Each entry of annotations[].errors:

Field	Type	Meaning
`span`	string	The exact phrase containing the error
`tags`	list of strings	Error code(s) for this span (`C1`–`C3` for contradiction, `N1`–`N3` for NEI)
`correction`	string	Corrected wording for the span; empty for NEI (nothing to replace with — the claim is unverifiable, not wrong)

Each entry of annotations[].2d_box:

Field	Type	Meaning
`page_id`	int	1-indexed PDF page number
`page_name`	string	Human-readable page label
`image_id`	string	Page image identifier
`coords`	`[int, int, int, int]`	Bounding box pixel coordinates `[x1, y1, x2, y2]`
`img_size`	`[int, int]`	`[width, height]` of the page image the coords are relative to

variant, target_label, question_valid, invalid_rationale, and source_pdf are omitted — unused downstream (question_valid is always true by construction: only pairs that passed question validation and were then annotated ever reach this dataset), and source_pdf was purely derivable from level == "full_paper".

License

Annotations (labels, rationale, error codes, dependency structure) are released under CC-BY-4.0. Source PDFs are from openly-accessible papers (NeurIPS, ICLR, ICCV, and similar venues); [TODO: confirm per-paper license terms before publishing -- this card asserts open accessibility, not a verified redistribution license for every included PDF].

Citation

[TODO: add citation once the paper has a venue/BibTeX entry.]

Downloads last month: 32

Number of rows:

Total file size:

16 MB