| --- |
| license: mit |
| pretty_name: Token Learning Spectrum Examples |
| tags: |
| - scaling-laws |
| - language-modeling |
| - token-loss |
| - interpretability |
| --- |
| |
| # Token Learning Spectrum Examples |
|
|
| This dataset hosts public `losses.npz` matrices for reproducing the figures in |
| the Token Learning Spectrum code release. |
|
|
| Each `losses.npz` follows the schema documented in the GitHub repository: |
|
|
| ```text |
| axis_values: float array [K] |
| loss_matrix: float array [N, K] |
| axis_name: string array [1] |
| sample_id, token_pos, and metadata arrays: optional arrays [N] |
| ``` |
|
|
| Files are listed in `manifest.yaml` with shapes, byte sizes, and SHA256 |
| checksums. Download them with: |
|
|
| ```bash |
| python tools/download_examples.py --repo-id applewpj/token-learning-spectrum-examples --all |
| ``` |
|
|
| The T/D/M-axis loss matrices are sanitized public analysis inputs. They do not |
| include private model architectures, weights, tokenizers, raw validation text, |
| training data composition, experiment registries, or checkpoint paths. |
|
|
| The synthetic Mano matrix is generated from the public synthetic arithmetic |
| pipeline adapted from PhysicsLM4. |
|
|
