A1 elicitation SFT LoRA over Qwen/Qwen2.5-32B base, trained 2026-08-04 on RunPod 4xH200 with LLaMA-Factory + DeepSpeed ZeRO-3.

Run B. Same targets as run A but lr 1e-4, 1 epoch, effective batch 8 (4xH200, accum 2). Also serves as epoch 1 of run C.

Qwen2.5-32B-elicit-A1-tablefix (lr 1e-4, tables trained 2 epochs) stops cleanly but takes an agent action in only 2 of 30 eval samples. The same adapter with its 4 table tensors deleted acts in 30 of 30. So training embed_tokens/lm_head appears to cost agent behaviour while fixing the end-of-turn terminator. These runs search for a setting that keeps both. Evidence: data/misalignment-eval/table-lora-debug/findings.md.

The acceptance test (action rate, stop rate, junk rate) has not been run on this adapter yet. Nothing here is known to work.

Serve with stop_token_ids=[151645, 151643]. Adapters containing token tables cannot be applied live by vLLM; merge first.