Ibisbill/olmo3-lit50b

olmo3_lit50b — story-filtered dolma3 literature partition, frozen 50B-token stream

A 50B-token stream built only from the common_crawl-literature-{0014..0019} partition of allenai/dolma3_mix-6T-1025-7B, keeping documents whose WebOrganizer/FormatClassifier p(Creative Writing) ≥ 0.9 (rule applied per unique text hash; upsampled copies inherit the decision; no dedup, so dolma's upsampling multiplicity is preserved). Documents are truncated to 512 tokens. Frozen with a seeded file shuffle (seed 20260729). This is the parent of olmo3_lit50b_shuf; it was not trained on directly.

Source: allenai/dolma3_mix-6T-1025-7B @ revision 2ca900fbe14e86c5c83d064d9f0882f1c0b8c05b · Tokenizer: allenai/dolma2-tokenizer (copy included under tokenizer/) · Created by: scripts/freeze_lit_stream.py @ git c805e86e1b01 on 2026-07-30. Ordering policy: frozen_prefix_stream_seeded_file_shuffle_filtered (shuffle_during_training: false; not the official OLMo-3 data order). Training runs that consumed this stream: none directly (parent stream only).

• Child stream (the one actually trained on): Ibisbill/olmo3-lit50b-shuf.
• Filter accounting is in filter_accounting.json (60.9M rows seen → 13.46M rows kept, 3.26M unique texts).

Format — the order is the data

Stored tokens	52,428,800,001 (= 52,428,800,000 training tokens + 1 look-ahead token)
Shards	2098 files shards/tokens_000000.u32 … tokens_002097.u32
Shard size	25,000,000 tokens (100,000,000 bytes) each; last shard 3,800,001 tokens
Element type	uint32, little-endian, raw (no header); docs separated by EOS id 100257
Vocab	dolma2, 100,278 raw ids, padded to 100,352 in the model

Shard i holds global token offsets [25,000,000·i, 25,000,000·(i+1)). shard_manifest.jsonl lists every shard with its start_token, num_tokens and sha256; manifest.json records provenance; verification.json (where present) records the end-to-end check performed after freezing.

Training read the stream strictly sequentially with global_tokens_per_update = 131,072 (128 sequences × 1,024 tokens per update, split across DDP ranks), 400,000 updates = 52,428,800,000 tokens. Hence update k (0-based) trained on tokens [131,072·k, 131,072·(k+1)), i.e. shards ⌊131,072·k / 25,000,000⌋ … ⌊(131,072·(k+1) − 1) / 25,000,000⌋. The target for position t is token t+1, which is why one extra token is stored.

import json, numpy as np

def tokens(root, start, n):
    """Return stream tokens [start, start+n) as a uint32 array."""
    shards = sorted((json.loads(l) for l in open(f"{root}/shard_manifest.jsonl")), key=lambda e: e["start_token"])
    out, need = [], n
    for e in shards:
        s0, s1 = e["start_token"], e["start_token"] + e["num_tokens"]
        if s1 <= start or need == 0: continue
        mm = np.memmap(f"{root}/{e['shard']}", dtype="<u4", mode="r")
        a = max(start, s0) - s0; b = min(a + need, e["num_tokens"])
        out.append(mm[a:b]); need -= b - a
        if need == 0: break
    return np.concatenate(out)

G = 131_072
k = 55_000                                  # what did the model see at update 55,000?
batch = tokens("path/to/olmo3_lit50b", G * k, G + 1)
x, y = batch[:-1].reshape(128, 1024), batch[1:].reshape(128, 1024)

Decode with tokenizers.Tokenizer.from_file("tokenizer/tokenizer.json") or transformers.AutoTokenizer.from_pretrained("tokenizer/").

Verifying a download

python - <<'EOF'
import json, hashlib
for l in open("shard_manifest.jsonl"):
    e = json.loads(l)
    h = hashlib.sha256(open(e["shard"], "rb").read()).hexdigest()
    assert h == e["sha256"], e["shard"]
print("all shards match")
EOF

License

The token stream is a derivative of allenai/dolma3_mix-6T-1025-7B and is redistributed under the same terms (ODC-BY 1.0); per-subset terms of the underlying sources apply as documented by the source dataset. The bundled tokenizer is allenai/dolma2-tokenizer (Apache-2.0).

Source machine path: /mnt/nlpgpu-io1/data/maggie/explain-m0de-collapse/replicate/data/olmo3_lit50b. Uploaded 2026-09-02.