The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.

climbmix-run

Streaming, packed-token derivative of karpathy/climbmix-400b-shuffle for continual pretraining of appvoid/picollama2.

Format

  • Source: karpathy/climbmix-400b-shuffle
  • Tokenizer: appvoid/picollama
  • Token dtype: little-endian uint16
  • Target file size: ~5,000,000,000 bytes
  • Files: shards/shard-XXXXXX.uint16.bin
  • Sidecars: shards/shard-XXXXXX.meta.json
  • Documents are separated with tokenizer EOS.
  • Shards close at tokenizer-batch boundaries, so an individual file may slightly exceed the target.
  • source_state_after in each sidecar is the Hugging Face IterableDataset cursor after that shard.
  • The dataset is generated sequentially and each source row is intentionally consumed once.

This is a derived/tokenized redistribution of the MIT-licensed ClimbMix shuffle.

Downloads last month
-