| --- |
| license: apache-2.0 |
| library_name: pytorch |
| tags: [dynamic-chunking, tokenization, hnet, byte-level, beetle-hnet, legacy-trainer] |
| --- |
| |
| # spec_E_s0 |
|
|
| Byte-level H-Net run from [beetle-hnet](https://github.com/suchirsalhan/beetle-hnet), trained |
| with the legacy trainer (`legacy/tier1/train.py`). Split out of the bundled repo |
| [`AdaptiveChunking/hnet-chunking-spec-82m`](https://huggingface.co/AdaptiveChunking/hnet-chunking-spec-82m) into the Beetle per-run layout. |
|
|
| - **`step-N` branches**: 21 retained checkpoints (0, 1, 2, 4, 8, 16, 32, 64, 128, 256, 1000, 2000 ...), |
| each with `model.safetensors` and `config.json` (`model_config`, training config and the |
| `tied_weights` map: restore with `sd[dst] = sd[src]` for every entry). |
| - **`main`**: the latest weights plus `final.json` and `log.jsonl`. |
| - The legacy trainer kept no optimiser state, so there is no `training_state/` here. FLORES+ |
| boundary probes for this run are in the dataset |
| [`AdaptiveChunking/hnet-chunking-probes`](https://huggingface.co/datasets/AdaptiveChunking/hnet-chunking-probes) under `spec_E_s0/`. |
|
|
