Occupying-Mars/glm-bfcl-tokenbender-replica-adapters

GLM BFCL tokenbender-replica adapters

PEFT LoRA adapters from the GLM BFCL recovery run in Occupying-Mars/prism-capability-extraction.

• source commit: 15e74d6
• base checkpoint: zai-org/glm-4-9b-chat, using glm_native prompt and target formatting
• recipe: r32/alpha64 rsLoRA, all-linear, policy KL + CE, grad_accum 8
• eval set: BFCL single-call, 1007 examples, normalized exact

folders

• r0/adapter/: PEFT adapter files
• r0/details/: train summary, round summary, leak audit/attribution metadata when available
• r1/adapter/: PEFT adapter files
• r1/details/: train summary, round summary, leak audit/attribution metadata when available
• r2/adapter/: PEFT adapter files
• r2/details/: train summary, round summary, leak audit/attribution metadata when available
• r3/adapter/: PEFT adapter files
• r3/details/: train summary, round summary, leak audit/attribution metadata when available
• datasets/r0_strict10k/: strict 10k training mix used for r0
• datasets/tokenbender_curricula/r1-r3/: generated repair curricula used for r1-r3
• datasets/tokenbender_curricula/branches/b001-b002/: paused branch curricula generated before branch search was stopped
• dataset_manifest.json: uploaded dataset file sizes and row counts

normalized exact

round	full r32	k99074	k148611	k198148	k247685	k297222
r0	456	283	358	400	393	405
r1	436	294	351	368	394	400
r2	434	290	358	380	394	399
r3	357	272	312	340	341	353

Base full normalized exact: 534/1007.

Current observation: r1/r2 near-miss rounds did not improve the main frontier over r0; r3 regressed on full and masked scores.