Occupying-Mars/issue49-k160-r32-len1024-adapter

BFCL v2 strict-mix k160 r32 LoRA adapter

LoRA adapter for Qwen/Qwen3-8B, trained for the circuit-shotting BFCL single-call experiments.

Run details:

• dataset: data/bfcl_strict_10k_mix/train.jsonl
• mask attribution: runs/issue49_bfcl_single_call/relp_full.npz
• mask size: 160000 MLP channels
• LoRA rank: 32
• LoRA alpha: 64
• max sequence length: 1024
• training output: runs/issue49_bfcl_single_call/lora_bfcl10k_topk160000_r32_balanced_len1024_ckpt

Eval on the 1007-row BFCL single-turn/single-call slice, with canonicalization prompt and normalized exact matching:

condition	score
unmasked merged model	683/1007 = 67.83%
k160 masked merged model	418/1007 = 41.51%

Included files:

• adapter_model.safetensors: PEFT LoRA adapter weights
• adapter_config.json: PEFT adapter config
• tokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer artifacts
• train_summary.json: training summary
• eval_masked_summary.json: k160 masked eval summary
• eval_unmasked_summary.json: unmasked eval summary