Datasets:
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
pretty_name: Token Learning Spectrum Examples
|
| 4 |
+
tags:
|
| 5 |
+
- scaling-laws
|
| 6 |
+
- language-modeling
|
| 7 |
+
- token-loss
|
| 8 |
+
- interpretability
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# Token Learning Spectrum Examples
|
| 12 |
+
|
| 13 |
+
This dataset hosts public `losses.npz` matrices for reproducing the figures in
|
| 14 |
+
the Token Learning Spectrum code release.
|
| 15 |
+
|
| 16 |
+
Each `losses.npz` follows the schema documented in the GitHub repository:
|
| 17 |
+
|
| 18 |
+
```text
|
| 19 |
+
axis_values: float array [K]
|
| 20 |
+
loss_matrix: float array [N, K]
|
| 21 |
+
axis_name: string array [1]
|
| 22 |
+
sample_id, token_pos, and metadata arrays: optional arrays [N]
|
| 23 |
+
```
|
| 24 |
+
|
| 25 |
+
Files are listed in `manifest.yaml` with shapes, byte sizes, and SHA256
|
| 26 |
+
checksums. Download them with:
|
| 27 |
+
|
| 28 |
+
```bash
|
| 29 |
+
python tools/download_examples.py --repo-id <hf-user-or-org>/token-learning-spectrum-examples --all
|
| 30 |
+
```
|
| 31 |
+
|
| 32 |
+
The T/D/M-axis loss matrices are sanitized public analysis inputs. They do not
|
| 33 |
+
include private model architectures, weights, tokenizers, raw validation text,
|
| 34 |
+
training data composition, experiment registries, or checkpoint paths.
|
| 35 |
+
|
| 36 |
+
The synthetic Mano matrix is generated from the public synthetic arithmetic
|
| 37 |
+
pipeline adapted from PhysicsLM4.
|