| --- |
| license: mit |
| base_model: microsoft/unixcoder-base-nine |
| tags: |
| - code |
| - ai-generated-code-detection |
| - classifier |
| library_name: transformers |
| --- |
| |
| # code-detection-confound checkpoints |
|
|
| Three fine-tuned **AI-generated-code detection** classifiers from the |
| [`AMindToThink/code-detection-confound`](https://github.com/AMindToThink/code-detection-confound) |
| research project. All three are cross-entropy-only (CE) fine-tunes of |
| [`microsoft/unixcoder-base-nine`](https://huggingface.co/microsoft/unixcoder-base-nine); |
| they differ only in training data. |
|
|
| | Subfolder | Training data | |
| |---|---| |
| | `unixcoder_dc_ce/` | DroidCollection-Python | |
| | `python_raw_ce/` | HMCorp / Python | |
| | `java_raw_ce/` | HMCorp / Java | |
|
|
| Each `model.bin` (~481 MB) is a **raw PyTorch `state_dict`** — no `config.json` or tokenizer |
| is bundled. Load it on top of the `microsoft/unixcoder-base-nine` architecture + tokenizer. |
| The exact training command (`scripts/18_train_cgs_amp.py … --model_name_or_path |
| microsoft/unixcoder-base-nine`), data provenance, and the classification head are documented |
| in the source repo's `REPRODUCE.md`. |
| |
| Backed up here during a machine migration (2026-07-02); see the source repo for full |
| reproduction details. |
| |
