JevAny-27B-SFT v0.2

JevAny-27B-SFT is the recommended general JevAny checkpoint. It is a rank 16 LoRA adapter plus a pointer head for Qwen/Qwen3.8-27B. It maps shared state and typed questions to option probabilities without decoding answer tokens.

Intended Use

Use this checkpoint for bounded classification, routing, ranking, tool choice, agent actions, and confidence-aware decisions. It accepts choice, noul, and ordinal score questions through the JevAny API. It is not a chat or chain-of-thought model.

Training

The adapter and pointer head were trained on 107,278 records spanning preferences, agent and tool decisions, hard and many-choice reasoning, classification, policy, and native image and video evidence. A separate calibration partition sets the inference temperature. The base weights remain frozen.

Evaluation

Evaluation	Result
v2 development accuracy	90.34%
v2 development NLL	0.265
transfer-v9 accuracy	82.41%
MMLU-Pro accuracy	73.0%
AI2D accuracy	86.0%
MMMU accuracy	68.0%

Development metrics exclude a 100-question single-class VideoFeedback slice. It exercises the native video path but is not a meaningful capability benchmark. AI2D and MMMU use native images through the backbone's vision path. Full measurements and the Jev comparison are in the repository release results.

Limits

The released path accepts text, JSON-renderable state, native images, and native video. 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. The trained text context envelope is 2,048 packed tokens. Calibration can shift under new data, so validate thresholds on a deployment-specific calibration set. The adapter requires the separately distributed Qwen base weights and JevAny runtime.