Publish the model training history: 119 checkpoints from the p3/p4/p5/stage4/noise runs
61f9ad0 verified | license: mit | |
| tags: | |
| - nrr | |
| - checkpoints | |
| - training-history | |
| # NRR model training history | |
| Every `.onnx` checkpoint the [NRR repository](https://github.com/SamurAI-Official/NRR) training runs produced - {0} files, {1:N1} MB - with their paths preserved, so a fetched checkpoint lands where that repository's tools and tests look for it: | |
| ```powershell | |
| hf download SamurAI-Official/NRR-checkpoints models/p5/p4_tjit_20261025.onnx --local-dir . | |
| ``` | |
| Why this repository exists: `models/` is not in the source repository's git history, so the only copy of | |
| these checkpoints was one machine's disk. The *released* model lives in | |
| [SamurAI-Official/NRR](https://huggingface.co/SamurAI-Official/NRR) instead; this is the history behind | |
| it, kept so that a comparison against the model a phase table names can be reproduced, and so the Unity | |
| verification project has the checkpoint its jitter tests select. | |
| What is deliberately **not** here, because "models/ is 3.3 GB" is misleading: | |
| | what | size | why it is not here | | |
| | --- | --- | --- | | |
| | run output (649 logs, `.report.json`, CSV) | 10.8 MB | working state; the repository's docs carry the measured summaries of those runs | | |
| | `models/training-data/*.npz` | 3.27 GB | captured frame packs: gitignored, regenerable by the capture scripts in `tools/`, and not models | | |
| The checkpoints are grouped by the phase that produced them - `models/p3`, `p4`, `p5`, `stage4`, `stage4b`, | |
| `noise`, `noise-warmup`, `frontier`, `phase3`, `phase4`, `preflip` - and named with the date of the run. |