qwen3-8b-bfcl-sft
Qwen3-8B fine-tuned (full-parameter SFT) on successful BFCL agent trajectories
generated by a Qwen3.6-27B teacher (no-rethink scaffold, train split, n=8
rollouts, deduplicated successful main-agent calls).
- Base model: Qwen/Qwen3-8B
- Data: 591 call-level samples (~22k supervised tokens), loss on assistant
targets only, prompts rendered serving-style with
enable_thinking=False
- Training: 1 epoch, lr 5e-6, bf16, max_len 16384
Intended for agentic function-calling research (BFCL-style multi-step tool use).
