Open-sourcing AstaBrief, the fast report-generation model in Asta
4 hours ago
- AstaBrief 8B is an open-weights model for generating cited scientific reports, trained via supervised fine-tuning and direct preference optimization.
- It focuses on data quality, using citation density filtering to improve evidence grounding and attribution.
- The model generates reports in a single pass, achieving 3.5× faster generation (51.1 seconds) compared to proprietary pipeline (178.5 seconds).
- Evaluation shows AstaBrief matches proprietary models on answer and citation quality, with human study favoring it on citation accuracy.
- Open weights enable institutions to run locally, and early Asta user feedback shows 84.2% positive rate with 29.1% repeat usage.
- Future work includes finer-grained preference learning, RAG-plus-RL approaches, and better evaluation of evidentiary scope.