OpenAster1-data-60B

This repository contains the continued-pretraining data used for the OpenAster1 60B-token stage.

The data is stored in Megatron-LM indexed dataset format (.bin / .idx) and is organized into 1B-token shards. The original recipe targets 60B tokens. The audited trainable total is 54,940,010,598 tokens because the final shard is an accepted short shard.

Contents

• shards/shard_000 to shards/shard_054: training shards.
• validation/bin: channel-level validation datasets.
• validation/jsonl: validation JSONL files.
• recipes: per-shard mixture recipes.
• reports: audit summaries for token counts and shard readiness.

Data Mixture

The continued-pretraining mixture contains the following channels:

• old_replay: replay from the 20B scratch mixture.
• web_general: DCLM/FineWeb/OpenCSG educational web tails.
• math_stem: FineMath, RedPajama arXiv, and math/STEM reference data.
• code: OPC FineWeb code, The Stack, and Codeforces-style code data.
• zh_multilingual: Chinese FineWeb educational data and SkyPile-ZH.
• knowledge_reference: Pile reference data and RedPajama C4/CommonCrawl slices.
• stepfun_chat: StepFun/Qwen-style chat data.

Format

Each training shard contains a Megatron-LM prefix such as:

shards/shard_000/bin/qwen3_2b_cpt60b_shard_000_text_document

Use the prefix without the .bin or .idx suffix as the Megatron --train-data-path entry.

The validation prefixes are:

validation/bin/code_text_document
validation/bin/knowledge_reference_text_document
validation/bin/math_stem_text_document
validation/bin/old_replay_text_document
validation/bin/stepfun_chat_text_document
validation/bin/web_general_text_document
validation/bin/zh_multilingual_text_document

Audit

See:

• reports/rebuild1b_shard_audit_summary.json
• reports/rebuild1b_shard_audit_summary.tsv

These files record shard readiness, indexed token counts, trainable token caps, document counts, and curriculum-stage assignment.

License

This dataset is a mixture assembled from public datasets. Individual examples retain the terms of their original sources. Users should verify downstream usage against the licenses and terms of the underlying datasets.