Update README: parquet now primary, bin archived to data_bin/

README.md

@@ -6,21 +6,58 @@ language:
 - en
 size_categories:
 - 1B<n<10B
+pretty_name: Stew-10B
+tags:
+- pretraining
+- gpt2
+- tokenized
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/*.parquet
 ---
 
 # Stew-10B
 
-10,000,000,000 GPT-2 tokens (`gpt2` fast tokenizer), pre-tokenized into raw
-`np.uint16` little-endian binary shards of 500,000,000 tokens
-(~1 GB) each. Every document is followed by `<|endoftext|>` (id 50256).
+**10,000,000,000 GPT-2 tokens** pre-tokenized and packaged for language-model
+pretraining. Available in **two formats** in this repo:
 
-| Source | Config | Share | Tokens |
-|---|---|---|---|
-| wikimedia/wikipedia | 20231101.en | 25% | 2,500,000,000 |
-| openbmb/Ultra-FineWeb | - | 50% | 5,000,000,000 |
-| mlfoundations/dclm-baseline-1.0 | - | 25% | 2,500,000,000 |
+| Format | Location | Size | Rows | Best for |
+|---|---|---|---|---|
+| **Parquet** (primary) | `data/*.parquet` | 14.55 GB | 4,882,800 | HF `datasets`, streaming, dataset viewer |
+| **Raw uint16 binary** | `data_bin/*.bin` | 18.63 GB | — | nanoGPT-style training, zero-copy `memmap` |
 
-## Usage
+Both formats contain the exact same token stream — the parquet files are a
+re-packaging of the `.bin` shards into fixed-length token windows of **2048 tokens** per row.
+
+## Mixture (10B tokens)
+
+| Source | Config | Share |
+|---|---|---|
+| `wikimedia/wikipedia` | `20231101.en` | 25% (2.5B) |
+| `openbmb/Ultra-FineWeb` | `en` | 50% (5.0B) |
+| `mlfoundations/dclm-baseline-1.0` | — | 25% (2.5B) |
+
+- Tokenizer: `gpt2` (fast / Rust BPE), vocab size 50257
+- Every document is terminated with `<|endoftext|>` (token id `50256`)
+- 20 shards, named `train-XXXXX-of-00020.{parquet,bin}`
+
+## Loading — Parquet (recommended)
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("wonder-ai/Stew-10B", split="train", streaming=True)
+for ex in ds:
+    tokens = ex["input_ids"]   # list of 2048 ints
+    break
+```
+
+Schema:
+- `input_ids`: `fixed_size_list<int32>[2048]` — one training window per row
+
+## Loading — Raw binary (fastest for training)
 
 ```python
 import numpy as np
@@ -28,11 +65,25 @@ from huggingface_hub import hf_hub_download
 
 path = hf_hub_download(
     "wonder-ai/Stew-10B",
-    "data/train-00000-of-00020.bin",
+    "data_bin/train-00000-of-00020.bin",
     repo_type="dataset",
 )
 tokens = np.memmap(path, dtype=np.uint16, mode="r")
 print(tokens.shape, tokens[:16])
 ```
 
-Each source retains the licensing of its upstream dataset.
+Each `.bin` is a flat little-endian `uint16` stream; the full set reconstructs
+to 10,000,000,000 tokens in order.
+
+## Reconstruction note
+
+The parquet files were generated from the `.bin` files by reshaping the token stream into 2048-token rows (any trailing tokens that did not fill a full row were dropped per shard).
+Document boundaries are preserved in both formats via the EOS token.
+
+## Licensing
+
+Each source retains the licensing of its upstream dataset. Please consult the
+original dataset cards before use:
+- Wikipedia: CC BY-SA 3.0 / GFDL
+- Ultra-FineWeb: see `openbmb/Ultra-FineWeb`
+- DCLM: see `mlfoundations/dclm-baseline-1.0`