| --- |
| language: en |
| license: apache-2.0 |
| library_name: pytorch |
| pipeline_tag: text-generation |
| tags: |
| - experimental |
| - checkpoints |
| --- |
| # Slayer149 training checkpoints |
|
|
| Automatic recovery archive for controlled data-continuation experiments from |
| [SlayerLab/Slayer149](https://huggingface.co/SlayerLab/Slayer149). |
| 148,910,738 parameters. Experimental checkpoints are not promoted releases or |
| verified top-three leaderboard entries. The original release remains unchanged. |
|
|
| `baseline/training-state.pt` and each `runs/<arm-seed>/checkpoint-<step>/training-state.pt` |
| contain weights, optimizer state, step, training configuration and provenance. |
| `extension/` contains the longer continuation only if the development selection |
| criteria pass. Steps are local to each training phase, not total lineage tokens. |
| Safetensors inference exports are included for the final pilot/extension checkpoints. |
| These use the accompanying custom PyTorch loader, not Transformers AutoModel. |
|
|
| `reports/` contains development measurements and experiment status. Development |
| proxies are not GLINT leaderboard results. Full GLINT confirmation, when available, |
| is explicitly named. Scores and ranking claims must be read with their protocol. |
| No training corpus or authentication credentials are uploaded. |
|
|
| Restore a trusted `training-state.pt` as `latest.pt` in a new run directory, |
| restore its matching config and exact tokenized data, then resume the trainer. |
| Source/tokenizer revisions and data hashes are recorded in the manifests; this |
| archive does not by itself contain the training datasets. |
|
|
