Kaon 3 CR-5C calibrated mean — GRPO step 50
Standalone BF16 full-weight merge; this is not an adapter-only repository.
- Base:
kaonai/kaon-c-gemma4-26b-v10.1@b17ba3fdebff342e303603225e1f0f1e6e606c16 - Source run/version:
cr5-mean-v11-order56-resume/v0-20260909-121159 - Optimizer step:
50 - GRPO LR / beta / seed:
1e-4/0.04/42 - Reward aggregation:
calibrated meanacross calibrated R/S/W margins - Sampling:
N18 → bottom3 + top3; strict three-way sign consensus - Adapter SHA-256:
4b9d971d5ede78091d452171db44f585925f047098aaeac8ada3349964844107 - Merge dtype:
bfloat16; saved/reloaded representative-logit parity:PASS - Status:
suggested unevaluated checkpoint; publication is not promotion authorization
See MERGE_AUDIT.json for source identity and verification details, and MANIFEST.sha256 for local file hashes.