The model is based on Qwen3-8B, initialized from AstaBrief-8B-SFT, and further trained on the AstaBrief_DPO_Mix dataset.
The model card says the preference-training set contains two reports for each real user query, with reports generated through multi-step Asta ScholarQA and single-step pipelines. Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini, GPT-4.1, DeepSeek-V3, and DeepSeek-R1 were used as backing models. GPT-4.1 and DeepSeek-R1 judged report pairs; Ai2 retained pairs where both agreed and reports 95% agreement between those judges and human preferences.
On the 100-question ScholarQA-CS2 test set, Ai2 reports an average score of 87.0, versus 83.7 for AstaBrief-8B-SFT and 77.3 for Qwen3-8B. Reported citation precision is 90.5 and citation recall 78.2. The source does not establish independent replication, performance outside its named benchmarks, or the reliability of reports when retrieval inputs are incomplete or incorrect.