README.md

---
license: other
license_name: cc-by-4.0-and-cc-by-sa-4.0
license_link: https://huggingface.co/datasets/nvidia/Nemotron-Math-v2
task_categories:
- text-generation
tags:
- math
- reasoning
- rl
size_categories:
- 10K<n<100K
---

# Nemotron-Math-v2-10k

10,000 machine-verifiable, mid-to-hard math problems subsampled from
[nvidia/Nemotron-Math-v2](https://huggingface.co/datasets/nvidia/Nemotron-Math-v2), each with all
stored gpt-oss-120b **high-reasoning, no-tool** solution traces. Built for RL training
(verifiable final answers) with a teacher-trace column for SFT distillation / analysis.

## How it was made (seed 42 throughout)

1. **Dedupe**: scanned all 5 splits (`high_part00-02`, `medium`, `low`), deduped rows by
   `md5(problem)` → 324,233 unique problems.
2. **Verifiable-answer filter**: kept problems whose `expected_answer` passes tier ≤ 2 of a
   whitelist filter (`answer_filter.py` in the scale `models` repo,
   `tmp/scalar-ppo/docs-ppo/scripts/`): tier 1 = plain number, tier 2 = short closed-form
   (whitelisted LaTeX commands, single-letter variables, ≤ 40 chars, no `\text`/`\sum`/
   matrices/relations). This removes answer shapes that `math_verify`-style graders can't
   check reliably (67% of raw Nemotron answers fail it).
3. **Difficulty buckets**: bucketed by `metadata.reason_high_no_tool.pass` — how many of the 8
   gpt-oss-120b attempts (high reasoning, no tools) were verified correct. Kept buckets
   **4, 5, 6, 7** (mid-to-hard band; excludes both never-solved and saturated problems).
   Tier≤2 pool sizes: {4: 7796, 5: 8483, 6: 9481, 7: 12744}.
4. **Sample**: 2,500 random problems per bucket = 10,000.
5. **Teacher traces**: for each problem, collected every stored row from the `high_part*`
   splits with `tools == []` (high-reasoning no-tool generations) and kept the **distinct**
   `messages` payloads (first-seen order) in `traces_high_no_tool`. Note the source release
   heavily duplicates payloads — a problem's no-tool row count tracks its pass statistics,
   but those rows carry only 1..pass distinct solutions (byte-identical repeats otherwise).
   Distinct-trace distribution over the 10k: 1 traces: 6818 · 2 traces: 172 · 3 traces: 379 · 4 traces: 738 · 5 traces: 731 · 6 traces: 643 · 7 traces: 458 · 8 traces: 11 · 9 traces: 10 · 10 traces: 9 · 11 traces: 9 · 12 traces: 7 · 13 traces: 4 · 14 traces: 8 · 15 traces: 2 · 18 traces: 1
   (9914/10000 problems have `n_traces <= pass_bucket`).
6. **Split**: shuffled, first 200 → `validation`, remaining 9,800 → `train`
   (validation bucket mix: {4: 38, 5: 58, 6: 44, 7: 60}).

## Trace caveats

- Traces were verified by NVIDIA's pipeline, not ours. In a 150-problem spot check with a
  strict `$`-wrapped `math_verify` against `expected_answer`, **~77% of distinct traces'
  final `\boxed{}` answers verify**; the rest are mostly verbose boxed formats
  (`\text{...}` prose, approximations like `n \approx 9.4456`) rather than wrong math.
  Re-filter per-trace if you need strict-format SFT targets.
- Every trace ends with a `\boxed{}` answer (0 missing in the spot check).

## Columns

| column | description |
|---|---|
| `problem` | problem statement (LaTeX) |
| `expected_answer` | verified final answer (majority-vote corrected where flagged) |
| `original_expected_answer` | pre-correction answer (differs only when `changed_answer_to_majority`) |
| `changed_answer_to_majority` | whether NVIDIA replaced the reference answer by majority vote |
| `pass_bucket` | gpt-oss-120b high-reasoning no-tool pass count (of 8) — 4..7 |
| `answer_tier` | 1 = plain number, 2 = closed-form expression |
| `data_source` | `aops` or `stackflow` |
| `license` | per-row source license (cc-by-4.0 for AoPS, cc-by-sa-4.0 for StackExchange) |
| `url`, `user_name`, `user_url` | StackExchange attribution (null for AoPS) |
| `metadata` | JSON: pass/count/accuracy for all 6 (reasoning × tool) teacher configs |
| `traces_high_no_tool` | list of JSON message arrays — all **distinct** stored high-reasoning no-tool teacher traces (`[user, assistant]`; `reasoning_content` = thinking, `content` = final answer with `\boxed{}`) |
| `n_traces` | `len(traces_high_no_tool)` (distinct payloads) |

## License / attribution

Derived from [nvidia/Nemotron-Math-v2](https://huggingface.co/datasets/nvidia/Nemotron-Math-v2)
(gpt-oss-120b generations). AoPS-sourced rows are CC-BY-4.0; StackExchange-sourced rows are
CC-BY-SA-4.0 with per-row attribution in `url`/`user_name`/`user_url`. See the per-row
`license` column.