🤖 Ai2 releases AstaBrief 8B weights for scientific reports with citations
A fine-tune of Qwen3-8B, trained via SFT and DPO without RL, writes a cited report in a single pass from a query and paper fragments. It was trained on 47,000 SFT examples from real ScholarQA queries and ~6,000 DPO pairs selected by GPT-4.1 and DeepSeek-R1 judges.
🌍 The open weights allow institutions to run review generation on their own infrastructure, including sensitive data. In 72% of pairwise comparisons, the model outperformed the multi-step Asta pipeline on Claude, and the report is prepared in 51 seconds versus 178.
👤 Weights are on Hugging Face: the 8B model runs locally via vLLM or transformers and compiles reports from its own PDFs (example in ai2-scholarqa-lib). A Fast mode is available without installation at asta.allen.ai. The model does not search on its own: fragments must be provided along with the question.
Source 1: https://allenai.org/blog/astabrief Source 2: https://huggingface.co/allenai/AstaBrief_8B