| --- |
| license: odc-by |
| language: |
| - en |
| task_categories: |
| - text-generation |
| pretty_name: HAGI tokenized FineWeb-Edu (SmolLM2) |
| tags: |
| - hagi |
| - pretraining |
| - tokenized |
| - fineweb-edu |
| - clifford-algebra |
| --- |
| |
| # HAGI - Tokenized HuggingFaceFW/fineweb-edu (SmolLM2-135M tokenizer) |
|
|
| Pre-tokenized token-id shards used to train the [HAGI](https://github.com/ShmidtS/HAGI) |
| Stage 0 baseline and the **Grade-Decomposed Recurrence** ablation (models A/B/C/D). A |
| **tokenized derivative** of [`HuggingFaceFW/fineweb-edu`](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) subset `sample-10BT`, |
| published so the exact training corpus loads identically in any environment (Colab, |
| Kaggle, local) with **no re-tokenization** and no Google Drive access. |
|
|
| ## Format - read before using |
|
|
| - **Files:** `shard_NNNNN.bin` - a flat little-endian **`uint16`** stream of token ids, |
| no header, documents separated by the tokenizer's EOS id. Read via `numpy.memmap`. |
| - **Tokenizer:** [`HuggingFaceTB/SmolLM2-135M`](https://huggingface.co/HuggingFaceTB/SmolLM2-135M) - **vocab 49,152** |
| (fits uint16). Shards tokenized with **any other tokenizer are incompatible**: the ids |
| would index the wrong embeddings (silent garbage), so do not mix sources. |
| - **Token count of a shard:** filesize / 2. |
| - **Held-out convention:** the **last shard** (`shard_00006.bin`) is reserved as validation; train on the rest. |
|
|
| ## Contents |
|
|
| - **Shards:** 7 |
| - **Total tokens:** ~630,891,027 |
|
|
| | Shard | Tokens | |
| |-------|--------| |
| | `shard_00000.bin` | 100,694,996 | |
| | `shard_00001.bin` | 100,686,108 | |
| | `shard_00002.bin` | 100,677,968 | |
| | `shard_00003.bin` | 100,324,700 | |
| | `shard_00004.bin` | 100,641,381 | |
| | `shard_00005.bin` | 100,534,502 | |
| | `shard_00006.bin` | 27,331,372 | |
|
|
| ## Load |
|
|
| ```python |
| from huggingface_hub import snapshot_download |
| path = snapshot_download(repo_id="NAME0x0/hagi-fineweb-edu-smollm2", repo_type="dataset", allow_patterns="*.bin") |
| # point training at the returned directory: |
| # python -m prototype.training.train --config configs/ablation_b.yaml --data <path> ... |
| ``` |
|
|
| `MemmapTokenDataset` (`prototype/data/dataset.py`) consumes this directory directly. |
|
|
| ## Models trained on this data |
|
|
| - Stage 0 baseline: https://huggingface.co/NAME0x0/hagi-stage0 |
| - Ablation: [`-a`](https://huggingface.co/NAME0x0/hagi-ablation-a) / [`-b`](https://huggingface.co/NAME0x0/hagi-ablation-b) / [`-c`](https://huggingface.co/NAME0x0/hagi-ablation-c) / [`-d`](https://huggingface.co/NAME0x0/hagi-ablation-d) |
|
|
| ## License & attribution |
|
|
| Derivative of HuggingFaceFW/fineweb-edu (sample-10BT), released under **ODC-By 1.0** - the upstream |
| license. Attribute FineWeb-Edu (HuggingFaceFW) on use. Tokenization adds no new content; |
| it only maps text to HuggingFaceTB/SmolLM2-135M ids. |
|
|
