Weights are published at loveCloud/AutoTCR under the filename AutoTCR.pth. This GitHub repository ships matched small assets in src/autotcr/assets/ and does not include the large checkpoint.
Download a complete offline bundle:
autotcr download-model --output-dir models/autotcr
autotcr validate-config --model-dir models/autotcr --device cpuThe directory will contain AutoTCR.pth, config.json, vocab.txt, tokenizer_config.json, inference_config.yaml, model_source.json and checksums.json. Use the whole directory for offline deployment.
The official release is pinned to commit 909bcd38137e73b8a94e709f3758b5d1a8180ebd. Checkpoint size is 230268116 bytes and SHA-256 is:
8949531ceee16f9919b90165d388460c3700d6971958866894ab893076bd58c3
The vocabulary contains 103 entries, preserving author-supplied order. Special IDs: PAD $=0, MASK .=1, UNK ?=2, SEP |=3, CLS *=4. Fixed inference length is32 and pooling is first-last-avg.
The full checkpoint includes BERT and the classifier, so a second pretrained weight file is not needed. Historical bundles containing 10k.pth are accepted by from_model_dir. Existing different model assets are not overwritten by the official download command.
Large weights and caches are ignored in this code repository. The author should separately state model-weight usage terms on the Hugging Face model card.