---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- math
- reasoning
- chain-of-thought
- gpt-oss
- rlvr
size_categories:
- n<1K
pretty_name: Math Reasoning (sample)
configs:
- config_name: sft
  default: true
  data_files:
  - split: train
    path: sft.jsonl
- config_name: labeled
  data_files:
  - split: train
    path: labeled.jsonl
---

# Math Reasoning, sample dataset

A small, illustrative sample for fine-tuning a model to **solve math problems step by
step and give a checkable final answer**. It accompanies the cookbook
[Fine-tune gpt-oss for math reasoning with Halo](https://github.com/whitecircle/halo).

This is a **format-illustration sample (48 rows)**, not a production training set. It
shows the exact record shape the recipe expects; swap in a larger corpus at scale.

## Configs

| Config | File | Schema | Used by |
|---|---|---|---|
| `sft` (default) | `sft.jsonl` | `{"prompt": [system, user, assistant]}` | supervised fine-tuning |
| `labeled` | `labeled.jsonl` | `{"question": str, "answer": str}` | online-GRPO verifiable reward |

The `sft` split teaches the format, reason through the problem then state the final
answer wrapped in `\boxed{...}` so it can be parsed with one regex. The `labeled` split
carries the same problems with only the gold answer, for the reinforcement-learning stage
that rewards correct answers directly.

## Record example (`sft`)

```json
{
  "prompt": [
    {"role": "system", "content": "You are a math tutor. Solve the problem step by step, then give the final answer on the last line as \\boxed{...}."},
    {"role": "user", "content": "A right triangle has legs 6 and 8. Find the hypotenuse."},
    {"role": "assistant", "content": "By the Pythagorean theorem, h = sqrt(6^2 + 8^2) = sqrt(100) = 10.\n\nThe answer is \\boxed{10}."}
  ]
}
```

## Composition

48 problems spanning arithmetic, fractions and percentages, linear and quadratic
equations, geometry, number theory, combinatorics and probability, sequences, and word
problems (rates, mixtures, work, compound interest). Answers are integers, simple
fractions, or a small radical, chosen so an exact-match grader can score them cleanly.

## Intended use

Supervised fine-tuning and RLVR for step-by-step math reasoning, as in the linked
cookbook. For a real run, scale to a few thousand problems with worked solutions and
raise the completion budget so the model has room to reason.