JevAny-27B-RLCR v0.2

JevAny-27B-RLCR is an experimental continuation of JevAny-27B-SFT using reinforcement learning with calibration rewards. It retains the same rank 16 LoRA and pointer-head architecture.

Intended Use

Use this checkpoint to study calibration-reward training for bounded decisions. JevAny-27B-SFT remains the recommended general checkpoint because RLCR did not improve overall development or transfer accuracy.

Training

The 40,000-record RL mixture emphasizes hard reasoning, many-choice questions, agent actions, preferences, mathematical and medical decisions, and image-derived and video-derived cases. The objective combines group-relative calibration reward with a supervised anchor. It is decision-only RL, not token-level GRPO, and it generates neither reasoning traces nor confidence tokens.

Evaluation

Evaluation	Result
v2 development accuracy	89.74%
v2 development NLL	0.260
transfer-v9 accuracy	82.31%
MMLU-Pro accuracy	73.5%
AI2D accuracy	87.0%
MMMU accuracy	63.0%

Development metrics exclude a 100-question single-class VideoFeedback slice. Against SFT, transfer accuracy changed by -0.10 percentage points, with 3 fixes and 4 regressions. The paired 95% bootstrap interval is [-0.58, 0.39] points. Full measurements are in the repository release results.

Limits

The small development NLL gain did not transfer consistently after independent calibration. Native multimodal requests currently contain one isolated question and use a bounded visual token budget. HTTP media is operator opt-in through a controlled local root; network media URLs are rejected. Confidence requires deployment-specific validation. The adapter requires the separately distributed Qwen base weights and JevAny runtime.