Repository navigation
This public release provides the frozen Qwen classification and joint-adaptation LoRA checkpoints, supervised readout weights, and standardizers for the primary Qwen experiments.
- Download model-assets-v1.zip (23,173,918 bytes; approximately 23 MB).
- SHA-256:
7d49ca8741d93180b62c19532c24af759d3157eabbde369374f336fd2e7d2811. - The independent Hugging Face evaluation Dataset provides anonymized human annotations and adjudication labels, split/contact manifests, cached features, frozen predictions, and supporting statistics. It contains no model checkpoints.
- Training, inference, and evaluation reconstruction scripts are available in this GitHub repository.
The asset bundle contains neither Qwen base-model weights nor original COCO images. Obtain these from their official sources. All third-party models and data remain subject to their original licenses and terms; the repository's MIT license applies to the original code, not all third-party materials.
The published archive is unchanged to preserve its checksum. Its bundled README retains historical preparation-time visibility wording; this release and its download are public. See the current repository README and Dataset Card for resource locations, access instructions, and component licensing. CPU checkpoint replay and prediction-table reconstruction are supported; GPU training and inference were not rerun for this release.