| AutoJev 27B curated SFT checkpoint | |
| Derived from Qwen/Qwen3.8-27B, revision | |
| 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0, under Apache License 2.0. | |
| Modifications: the language-generation output layer is replaced with a | |
| 255-option decision readout; retained backbone, vision and readout weights | |
| were updated through supervised fine-tuning. A scalar temperature fitted | |
| on a separate calibration split is saved in decision_config.json. | |
| Training and serving code under source/ is separately MIT licensed. | |
| No affiliation with or endorsement by Qwen or TypeSafe is implied. | |