| --- |
| language: |
| - en |
| license: other |
| library_name: transformers |
| pipeline_tag: fill-mask |
| tags: |
| - babylm |
| - strict-small |
| - research-release |
| --- |
| |
| # Structured Direct-Sum randomized-prior control |
|
|
| This repository preserves an already-trained checkpoint from the controlled |
| BabyLM research tournament. No training or evaluation was run for this release. |
|
|
| ## Selected revision |
|
|
| `main` is identical to `chck_100M`, selected from the existing local tournament |
| record. Other revisions, when present, are archived checkpoints rather than new |
| experiments. |
|
|
| ## Existing evaluation record |
|
|
| | Evaluation | Score | |
| |---|---:| |
| | BLiMP | 66.02 | |
| | Supp | 59.60 | |
| | EWoK | 49.64 | |
| | ET | 20.75 | |
| | COMPS | 52.47 | |
| | GlobalPIQA | 37.62 | |
|
|
| The canonical code is maintained at |
| [miguelcsx/tolm](https://github.com/miguelcsx/tolm). Remote code is required to |
| load this custom Transformers model; review `tolm.py` before use. |
|
|
