license: cc-by-4.0
language:
- en
tags:
- math
- reasoning
- sft
- distillation
size_categories:
- 100K<n<1M
OpenMathReasoning_Subset_8k
The cap-8192 sibling of bevangelista/OpenMathReasoning_Subset (which caps at 4096). Same source, same filters, same one-shortest-trace-per-problem selection — only the token cap differs — so the two datasets form a controlled length-prior axis for compute-matched SFT ablations on gemma-3-4b. This dataset is a strict superset of the 4k subset: its first 158,638-row file contains every 4k row byte-identically, plus 87,034 problems whose shortest trace lands in 4096–8192 tokens (522.9M tokens of long traces).
Attribution
Derived from nvidia/OpenMathReasoning, licensed CC-BY-4.0 © NVIDIA Corporation. Redistributed under CC-BY-4.0 with attribution, as its terms require.
What was done
From all 144 cot parquet files, keep rows where: teacher generation_model == DeepSeek-R1;
inference_mode ∈ (null, cot); a parseable, non-prose answer (no newline, ≤120 chars); a real
chain-of-thought; n_tokens ≥ 256; and difficulty gate pass_rate_72b_tir ≤ 0.95 (null pass_rate
KEPT). Deduped to one shortest trace per unique problem (blake2b(problem)), then capped at
n_tokens ≤ 8192. text is the fully formatted gemma-3 training string; n_tokens is its length
under the gemma-3-4b tokenizer. "Shortest" is shortest-by-characters (selection runs before
tokenization), matching the 4k subset exactly.
Columns
| column | type | meaning |
|---|---|---|
text |
string | formatted training example (gemma-3 chat template, <think>…</think> + answer) |
n_tokens |
int32 | token length of text under gemma-3-4b tokenizer |
pass_rate |
float32 | 72B pass rate (higher = easier); null when the source had none |
band |
string | always core here (≤8192 tokens) |
Stats
158,638 problems · 714.5M tokens · mean 4,504 / median 4,406 tokens · mean pass_rate 0.527 (present on 52% of rows).