| --- |
| license: other |
| language: |
| - en |
| task_categories: |
| - text-generation |
| tags: |
| - gpt2 |
| - tokenized |
| - long-context |
| - pretraining |
| size_categories: |
| - 10B<n<100B |
| --- |
| |
| # BluScriptCP tokenized token buckets (b1, b2, b3) |
|
|
| Pre-tokenized GPT-2 data for long-context training in the BluScriptCP / Context |
| Parallelism project. These are the three mid-length buckets of a token-bucketed, |
| deduplicated CommonCrawl-derived corpus (`new_full_2025_51+47`, dclm+betr). |
|
|
| Documents are bucketed by GPT-2 token length into half-open ranges `[lo, hi)`. |
| Only b1, b2 and b3 are published here. |
|
|
| | bucket | token range | shards | documents | GPT-2 tokens | size | |
| |---|---|---|---|---|---| |
| | `b1_4096_8192` | [4,096, 8,192) | 96 | 2,571,614 | 13,977,775,051 | 28.0 GB | |
| | `b2_8192_16384` | [8,192, 16,384) | 96 | 634,512 | 6,990,738,341 | 14.0 GB | |
| | `b3_16384_32768` | [16,384, 32,768) | 96 | 142,657 | 3,027,863,374 | 6.1 GB | |
| | **total** | | **288** | **3,348,783** | **23,996,376,766** | **48.1 GB** | |
|
|
| ## Format |
|
|
| Tokenizer is `tiktoken` encoding `gpt2` (vocab 50,257), applied with |
| `encode_ordinary`, which treats `<|endoftext|>`-like text as ordinary tokens. |
|
|
| Each of the 288 shards has three files under `<bucket>/`: |
|
|
| | file | dtype | contents | |
| |---|---|---| |
| | `part_XX.bin` | `uint16` | GPT-2 token ids, documents concatenated, no separator | |
| | `part_XX.idx.npy` | `int64` | cumulative offsets, length `ndocs + 1` | |
| | `part_XX.meta.json` | - | per-shard document count, token count, verification counters | |
|
|
| Token ids fit `uint16` losslessly since the GPT-2 vocabulary is 50,257 < 65,536. |
|
|
| **No EOT separator is inserted.** Document boundaries are carried exactly by the |
| index, so a packing step must insert its own delimiter if one is needed. |
|
|
| ## Usage |
|
|
| ```python |
| import numpy as np |
| |
| stem = "b1_4096_8192/part_00" |
| idx = np.load(stem + ".idx.npy") # int64, ndocs+1 |
| tok = np.memmap(stem + ".bin", dtype=np.uint16, mode="r") # uint16 stream |
| |
| n_docs = len(idx) - 1 |
| doc_0 = tok[idx[0]:idx[1]] # document 0 as a uint16 array of token ids |
| |
| import tiktoken |
| enc = tiktoken.get_encoding("gpt2") |
| print(enc.decode(doc_0.tolist())[:200]) |
| ``` |
|
|
| To iterate every document in a shard: |
|
|
| ```python |
| for i in range(len(idx) - 1): |
| doc = tok[idx[i]:idx[i + 1]] |
| ``` |
|
|
| ## Verification |
|
|
| Every document in the source corpus carries a pre-computed `gpt2_tokens` count. |
| Each document's re-tokenized length was compared against it: |
|
|
| - **0 count mismatches across all 3,348,783 documents** in these three buckets |
| (and 0 across all 74,643,792 documents of the full 7-bucket corpus). |
| - Document and token totals are an exact match to the corpus `summary.tsv`. |
| - Index integrity checked per shard: `len(bin) == idx[-1]`, `idx[0] == 0`, `idx` |
| strictly increasing, `max(token_id) < 50257`. |
| - Exact text roundtrip: `enc.decode(bin[idx[i]:idx[i+1]])` is byte-identical to the |
| source `content` field on sampled documents from every bucket. |
|
|
| ## Not included |
|
|
| The other four buckets of the corpus exist but are not published here: |
| `b0_lt_4096` (71.25 M docs, 73.59 B tokens), `b4_32768_65536`, `b5_65536_131072`, |
| `b6_gte_131072`. |
|
|
| ## Provenance |
|
|
| Derived from CommonCrawl via the DCLM and BETR filtering pipelines, then |
| deduplicated. Records retain their original scores and metadata upstream; only the |
| token ids are published here. Respect the licensing and terms of the underlying |
| CommonCrawl data. |
|
|
