README.md

metadata

license: odc-by
language:
  - en
task_categories:
  - text-generation
pretty_name: lilm-pretrain
size_categories:
  - 100K<n<1M

lilm-pretrain

A static mixed pretraining dataset curated for a sub-200M language model.

Repository: glouriousgautam/lilm-pretrain

Source mix

• finepdfs: codelion/finepdfs-1B config default, target 500,000,000 tokens
• dclm_baseline: codelion/dclm-baseline-1B config default, target 300,000,000 tokens
• fineweb: HuggingFaceFW/fineweb config default, target 40,000,000 tokens
• fineweb_edu: codelion/fineweb-edu-1B config default, target 200,000,000 tokens
• finemath_4plus: HuggingFaceTB/finemath config finemath-4plus, target 10,000,000 tokens

Curation method

Each source is streamed and sampled with reservoir sampling using seed 20260604. The sampled source shards are merged and deterministically shuffled into Parquet shards.

The token_count column is based on the source token_count field when available and otherwise on HuggingFaceTB/SmolLM2-135M tokenization.

Produced dataset

• Rows: 574,501
• Estimated tokens: 1,050,011,322
• Parquet shards: 23

Per-source result

• finepdfs: 89,924 rows, 500,003,682 tokens
• dclm_baseline: 229,952 rows, 300,001,747 tokens
• fineweb: 57,446 rows, 40,004,125 tokens
• fineweb_edu: 190,934 rows, 200,001,086 tokens
• finemath_4plus: 6,245 rows, 10,000,682 tokens

Columns

• text: training text
• source: short source alias
• source_dataset: Hugging Face source dataset id
• source_config: source config/subset
• source_split: source split
• source_row_id: source identifier or stable hash
• token_count: estimated token count
• sample_seed: sampling seed
• source_seen_index: row position seen during streaming