ashcash15/qwen2.5-7b-search

Roll back to v1 (20-trace adapter): v2 (53-trace, 2ep) overfit and collapsed tool-calling on ~47% of eval questions

Stage 1: adapter retrained on 53 traces (2.65x Stage 0)

Model card: Stage 0 training details, eval results, usage

LoRA adapter: Stage 0 agentic-search SFT (H100, 20 traces, 2 epochs)

Upload README.md with huggingface_hub

initial commit