allenai/AstaBrief_8B
allenai/AstaBrief_8B is an 8-billion parameter language model developed by Allen Institute for AI (AI2), fine-tuned from Qwen3-8B. It specializes in generating cited reports from research questions and scientific literature excerpts, utilizing supervised fine-tuning and offline direct preference optimization. With a 32K context length, it excels at synthesizing information into structured scientific reports.
Loading preview...
AstaBrief-8B: Scientific Report Generation
AstaBrief-8B, developed by Allen Institute for AI (AI2), is an 8-billion parameter model fine-tuned from Qwen3-8B. Its core function is to transform a research question and provided scientific literature excerpts into a cited report. This is achieved through a combination of supervised fine-tuning (SFT) and offline direct preference optimization (DPO).
Key Capabilities
- Cited Report Generation: Produces structured reports with citations based on user queries and scientific text.
- Preference Optimization: Fine-tuned using the AstaBrief_DPO_Mix dataset, which includes real user queries and preference judgments from LLM judges (GPT-4.1 and DeepSeek-R1) aligned with human preferences.
- Enhanced Performance: Demonstrates improved performance over its base Qwen3-8B model and the AstaBrief-8B-SFT checkpoint on the ScholarQA-CS2 test set, a benchmark for computer science research questions.
- Competitive Benchmarking: Achieves competitive results against models like Asta ScholarQA and DR-Tulu on various benchmarks, including SQA-CS2 and DeepScholarBench.
Good For
- Academic Research: Generating concise, cited summaries or reports from scientific literature.
- Information Synthesis: Automating the process of compiling information from multiple scientific sources into a coherent document.
- Research and Educational Use: Intended for applications aligned with AI2's Responsible Use Guidelines, particularly in academic and research settings.