Datasetpublic

Yujivus/nanochat-climbmix-arithmetic-base10

nanochat ClimbMix + Base-10 Arithmetic This dataset contains the first 170 shuffled ClimbMix training shards used by nanochat's speedrun. The deterministic base-10 arithmetic corpus is mixed into shards 00000..00149; the final 20 train shards are unchanged web-only padding. The original validation shard (shard_06542.parquet) is also copied unchanged. Arithmetic corpus Family Examples a + b = c (all ordered pairs 0..2000, two exposures) 8,008,002 a + b… See the full description on the dataset page: https://huggingface.co/datasets/Yujivus/nanochat-climbmix-arithmetic-base10.

sourceHugging Facecc-by-nc-4.0updated 1mo agoView on Hugging Face

0likes3.7kdownloads

Dataset Card

nanochat ClimbMix + Base-10 Arithmetic

This dataset contains the first 170 shuffled ClimbMix training shards used by nanochat's speedrun. The deterministic base-10 arithmetic corpus is mixed into shards 00000..00149; the final 20 train shards are unchanged web-only padding. The original validation shard (shard_06542.parquet) is also copied unchanged.

Arithmetic corpus

Family	Examples
`a + b = c` (all ordered pairs `0..2000`, two exposures)	8,008,002
`a + b + c = d` (1M unique ordered triples from `0..200`)	1,000,000
`a - b = c` (`0 <= b <= a <= 2000`)	500,000
`a * b = c` (`a=0..2000`, `b=1..20`)	40,020
`c / b = a` (exact inverses)	40,020
`a % m = r` (`a=0..2000`, `m=2..100`)	198,099
Total equations	9,786,141

Every arithmetic document contains 16 newline-separated equations (the final document contains fewer). This gives 611,634 arithmetic documents, distributed as evenly as possible across the first 150 training shards and interleaved throughout the original web documents. Arithmetic examples are globally shuffled with the fixed seed 42. arithmetic_manifest.json fully specifies the generator.

Provenance

—nanochat: https://github.com/karpathy/nanochat
—Source mirror: https://huggingface.co/datasets/Yujivus/nanochat-climbmix-170
—Karpathy's shuffled packaging: https://huggingface.co/datasets/karpathy/climbmix-400b-shuffle
—Pinned source revision: 863e849617372d74b01917d80d46afbea54660f9
—Upstream dataset: https://huggingface.co/datasets/nvidia/Nemotron-ClimbMix

The ClimbMix text is not authored here. The upstream NVIDIA dataset card lists CC BY-NC 4.0, so this mixed dataset is conservatively marked the same. Users must review and comply with all upstream terms.