YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Qwen2.5-32B-elicit-A1-tablefix
STATUS: CLEAN — strict junk test 2026-08-02: 0/120 trailing, 0/120 foreign-script, 118/120 clean stops (2 legitimately long answers hit the 400-token cap).
A1 elicitation SFT LoRA over Qwen/Qwen2.5-32B BASE: r64/alpha128/dropout 0, assistant-only loss, 13k A1 mix, 2 epochs, lr 1e-4 cosine + 3% warmup, cutoff 8192, neat_packing, 8xH100 ZeRO-3 (LLaMA-Factory).
THE FIX vs earlier contaminated base-start runs: LoRA targets include embed_tokens and lm_head. Qwen2.5 BASE ships identical untrained embedding/lm_head rows for the ChatML specials (QwenLM/Qwen3#1064), so a linear-only LoRA cannot make <|im_end|> selectable — per-row table deltas break that tie. Verified: P(<|im_end|>)=0.994 at completion points post-merge.
SERVING: vLLM cannot apply this adapter live (table LoRAs) — MERGE first: PeftModel.from_pretrained(base, adapter).merge_and_unload(), save, serve the merged model. Also pass stop_token_ids=[151645,151643] per request: the base generation_config only lists <|endoftext|> and can defeat server-side overrides.