jev-ettin-400m-20261013-v0

Private checkpoint for the JEV encoder decision model.

• backbone/: pretrained-compatible encoder configuration and fine-tuned Safetensors weights.
• tokenizer/: tokenizer files required to encode inputs.
• head.safetensors: trained candidate scoring head parameters.
• decision.json: model architecture and input-encoding configuration.
• calibration/calibration.json: learned temperature and the corresponding model fingerprint.
• upload-manifest.json: model file sizes and checksums.

The repository root is both the selected best and final checkpoint.

These are custom PyTorch decision-model artifacts. Load them using DecisionModel.load or EncoderEngine from your JEV project. The root calibration file is compatible with EncoderEngine(artifact=..., calibration=...); it applies only to the root checkpoint. A distinct final checkpoint must be used without this temperature calibration. This model format is not a standard AutoModelForSequenceClassification checkpoint.