Datasets:

allenai
/

AstaBrief_SFT_Mix

id string	source string	messages list	citation_density_sentences float64
b60f36a4-ccd9-5492-9e7e-574bdf87aaaf	sqatrace_gpt-4.1-2025-04-14_trace0	[ { "role": "user", "content": "\nA user issued a query and a set of research papers were provided with salient content.\nThe user query was: find me Explainable Agents and Robots surevy papers\n\nI will provide you with a list of chosen quotes from these papers that are relevant to the user query. \nYour job...	0.414894
a18a94bc-0c9a-59a4-8cb3-7caf2a247c83	sqatrace_gpt-4.1-2025-04-14_trace0	[ { "role": "user", "content": "\nA user issued a query and a set of research papers were provided with salient content.\nThe user query was: cutting edge machine learning methods for rr lyrae variable stars\n\nI will provide you with a list of chosen quotes from these papers that are relevant to the user que...	0.368421
3adba818-3668-5016-92f3-85035d4fbec8	sqatrace_azure/o3-2025-04-16_trace0	[{"role":"user","content":"\nA user issued a query and a set of research papers were provided with s(...TRUNCATED)	0.367089
6ee1f0cb-a746-5750-aa30-1a8ee3c523a8	sqatrace_azure/o3-2025-04-16_trace0	[{"role":"user","content":"\nA user issued a query and a set of research papers were provided with s(...TRUNCATED)	0.266667
9319d958-0105-5abe-84bf-ff3430404b28	sqatrace_azure/o4-mini-2025-04-16_trace0	[{"role":"user","content":"\nA user issued a query and a set of research papers were provided with s(...TRUNCATED)	0.345238
1cab5b5f-3b58-5717-ad4d-9117417e66a1	sqatrace_azure/o4-mini-2025-04-16_trace0	[{"role":"user","content":"\nA user issued a query and a set of research papers were provided with s(...TRUNCATED)	0.282353
4f91c2f0-6432-5b1e-81b0-0806e0085e3a	sqatrace_claude-3-5-sonnet-20241022_trace0	[{"role":"user","content":"\nA user issued a query and a set of research papers were provided with s(...TRUNCATED)	0.389474
004e7247-07f6-59ab-85b4-fae04f08a388	sqatrace_gpt-4.1-2025-04-14_trace0	[{"role":"user","content":"\nA user issued a query and a set of research papers were provided with s(...TRUNCATED)	0.494505
f1d17caa-b5b0-5dbe-bd67-c834aa7019f4	sqatrace_gpt-4.1-2025-04-14_trace0	[{"role":"user","content":"\nA user issued a query and a set of research papers were provided with s(...TRUNCATED)	0.402985
e4b642c1-8a1c-5bde-92bc-94da37b93a84	sqatrace_gpt-4.1-2025-04-14_trace0	[{"role":"user","content":"\nA user issued a query and a set of research papers were provided with s(...TRUNCATED)	0.362245

End of preview.

AstaBrief SFT Data

This dataset contains SFT training data for AstaBrief, consisting of 39.5k real user query prompts and final cited reports.

User Query Curation

User queries included in this dataset are sourced from real queries issued to Ai2's OpenScholar and Asta ScholarQA systems until June 2025. We only retain queries submitted by users who opted-in to data sharing. First, we strip out beta-tester and bot traffic and drop queries that are too short to be meaningful. The remaining queries go through a second filtering pass with GPT-4o, using this filtering prompt which focuses on dropping non-English queries, non-scientific requests, and prompts containing personal medical information.

We are left with a pool of 90k queries at the end of these filtering processes.

Report Generation

For each query, we generate a full report using the multi-step ScholarQA pipeline behind Asta's report generation functionality. The pipeline retrieves relevant literature, organizes the material into sections, and synthesizes the evidence into a cited report. We drew on a mix of proprietary models to back this pipeline: Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini, and GPT-4.1. After additional quality filtering (e.g., dropping reports with very few references), this yielded 47K usable training examples after setting some aside for DPO data generation.

Report Filtering

Our first SFT run improved overall content quality, but still lagged behind a Claude-powered report generation pipeline on answer precision and citation quality. To further improve citation quality, we measured the citation density of reports (i.e., the share of statements that had at least one citation) and filtered out reports with density<0.25, leaving us with 39.5k examples. Using this subset for SFT training led to a much higher-scoring SFT checkpoint, see the SFT checkpoint's model card for evaluation details.

License

This dataset is licensed under CC BY-NC 4.0. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines. The data contains synthetic outputs generated by third party models, which are subject to the terms and conditions of the respective model provider.