issue51-r32-more-online-codex-repro-551-full

Full merged model and receipts for a direct Lium A100 reproduction of the issue #51 r32_more_online BFCL k160 masked LoRA configuration.

Result

• Eval slice: BFCL v3 single-turn/single-call, 1007 rows
• Mask: relp_full.npz, topk=160000, zero-isolated eval
• Parent/init adapter: Occupying-Mars/issue49-k160-r32-len1024-adapter
• Normalized exact: 551/1007 = 0.547170
• Raw exact: 537/1007 = 0.533267
• Training updates: 125
• Training elapsed seconds: 3718.541

This run did not reproduce the remembered 562/1007 line; it landed at 551/1007 normalized exact. It is also slightly below the previously HF-backed gkd_aopd_hybrid_k160_v0 result at 553/1007.

Layout

• Repo root: full merged model saved from r32_more_online/merged
• adapter/: PEFT LoRA adapter saved from r32_more_online/adapter
• receipts/: train summary/log, eval summary/log, eval generations JSONL, and source pointer