README.md

metadata

license: odc-by
task_categories:
  - text-generation
language:
  - en

dolma3_6T_preconditioner_100k

Uniform random preconditioner sample from the Dolma3 6T deduplicated corpus.

Parameters

Parameter	Value
Type	uniform_random
Purpose	preconditioner_building
Target docs	100,000
Min tokens/doc	512
Seed	42

Counts

Metric	Value
Total documents	100,000
Total tokens	251,517,816
Unique source shards	38,277

Token distribution

Statistic	Value
Min	512
Max	468,399
Median	1,243
Mean	2,515.2

Format

JSONL files compressed with zstandard under data/part_NNN/. Each record contains the original Dolma document fields (id, text, metadata).

Sampling metadata files:

• working_sample_manifest.parquet - sampled doc_ids with shard paths
• sample_contract.json - sampling parameters and realized statistics