lora7b_rbonly_grid
3-way stop/deliver readout head (CONTINUE / DELIVER / ABORT) fine-tuned from Qwen2.5-7B-Instruct on trajectory-prefix windows
(dataset lllqaq/datal). Trainer + recipe: see split_info.json, train.log / lora.log. best/ = repo-held-out val-AUC-best
checkpoint, model/ (full) or adapter/ (LoRA) = final weights. scores_*.jsonl = per-window P(ABORT) on the frozen test file.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support