jaymanaryan/indra_pretraining_corpus_v1

INDRA raw-text pretraining corpus

This is a generated causal-LM pretraining / continued-pretraining corpus. Each example is an ordinary raw document in the text column; it does not contain SFT instruction/answer wrappers.

Counts

• Target tokens: 5,000,000
• Accepted tokens: 5,045,176
• Training documents: 12,339
• Validation documents: 635
• Exact token counting with target tokenizer: True

Target mixture

{
  "english": 0.25,
  "indic": 0.25,
  "world": 0.15,
  "math": 0.1,
  "science": 0.1,
  "code": 0.1,
  "structured": 0.05
}

Layout

• train/*.jsonl: raw documents for gradient updates
• validation/*.jsonl: held-out documents for validation loss only
• dataset_manifest.json: per-source scan counts, failures, licenses, and token counts

Each JSONL row has text, category, language, source, license, and token_count. The current Indra local JSONL loader reads the text field and ignores metadata during training.

Source/license warning

Source licenses vary. FineWeb-Edu is marked ODC-By 1.0; Wikimedia Wikipedia content is CC-BY-SA 3.0/GFDL; IndicCorpV2 is described as CC0; CodeParrot Clean retains per-record GitHub license labels and filters to common permissive licenses; OpenWebMath and Scientific Papers source terms must be reviewed before redistribution. Review and comply with source terms before public redistribution. This script does not grant a new license over the combined text.

See dataset_manifest.json for source-level success/failure and quota shortfalls.