Datasets:

zaydsyed
/

TactLM-Math-Reasoning-Data

---
pretty_name: TactLM Math Reasoning Data
language:
- en
license:
- mit
- apache-2.0
task_categories:
- question-answering
tags:
- math
- reasoning
- reinforcement-learning
configs:
- config_name: curated
  data_files:
  - split: train
    path: data/curated/train.jsonl
  - split: validation
    path: data/curated/validation.jsonl
  - split: test
    path: data/curated/test.jsonl
- config_name: fresh-multistep
  data_files:
  - split: test
    path: data/fresh-multistep/test.jsonl
---

# TactLM Math Reasoning Data

The questions used to train and evaluate [TactLM Math ProcessGRPO](https://huggingface.co/zaydsyed/TactLM-Math-ProcessGRPO-Qwen2.5-0.5B). The release has two subsets: the filtered training pool and its holdouts, and a separate generated test of multi-step arithmetic. Each row has a final answer. The step-credit rewards were computed during training and are **not** labels in this dataset.

| Subset | Split | Problems | Purpose |
| --- | --- | ---: | --- |
| `curated` | `train` | 8,478 | GRPO training |
| `curated` | `validation` | 380 | Checkpoint selection |
| `curated` | `test` | 510 | Held-out evaluation |
| `fresh-multistep` | `test` | 700 | Chained-operation evaluation |

## How the questions were selected

The curated subset draws from the **training splits** of [MATH](https://huggingface.co/datasets/EleutherAI/hendrycks_math), [GSM8K](https://huggingface.co/datasets/openai/gsm8k), and selected sources in [NuminaMath-CoT](https://huggingface.co/datasets/AI-MO/NuminaMath-CoT), plus problems made by this project's dependency-graph generator. The pipeline kept questions with a single numeric answer, removed duplicates and questions requiring figures, and screened for overlap with evaluation sets using digit-masked 13-grams. It then selected questions the starting Qwen2.5-0.5B-Instruct model answered correctly in **1–7 of 8 attempts**. A blind solver checked each retained source answer; disagreements were sent to a second solver. The pool was split by topic, and the training split was capped so no topic dominated it. A separate 200-question hand audit found no wrong answer keys.

The generated rows are an exception to the solvability filter: they have `source: generated` and `hits: null`. There are **847** in train and **40** in validation. They require 2–4 chained operations and include a distractor sentence.

The `fresh-multistep` test was generated **after training** with an unused seed. It contains 100 questions at each depth from 2 through 8 operations. The generator constructs an arithmetic dependency graph, computes the answer from it, and writes the question in English. These questions were used for the published model's evaluation, so future work should not treat this public release as a new blind test.

## Fields

`curated` rows have `id`, `problem`, `answer`, `source`, `subject`, `topic`, and `hits`. `answer` is a numeric string, sometimes a fraction. `hits` is the starting model's correct count out of eight attempts; it is null for generated rows. `subject` is populated for MATH rows. The `test` split is the exact 510-question set used by the evaluation harness.

`fresh-multistep` rows have `id`, `problem`, `answer`, and `ops`. `ops` is the number of arithmetic operations in the generated solution graph. The released file uses the generator's fixed seed `20261024`.

```python
from datasets import load_dataset

train = load_dataset("zaydsyed/TactLM-Math-Reasoning-Data", "curated", split="train")
fresh = load_dataset("zaydsyed/TactLM-Math-Reasoning-Data", "fresh-multistep", split="test")
```

While this repository is private, pass `token=True` to `load_dataset` after signing in.

## Limits and provenance

This set was chosen for one small model's training run, not as a general math benchmark. The solvability band excludes questions that model never solved. Topic tags and answer checks were model-assisted; the 200-question audit is a sample, not a guarantee that every key is correct. The generated questions follow templates and should not be mistaken for independent human-written problems. The held-out and fresh sets are now disclosed alongside the training data; report comparisons against them with that context.

Questions from upstream datasets retain their original source in the `source` field. Licensing and citations are listed in [SOURCE-LICENSES.md](SOURCE-LICENSES.md). The dataset combines MIT-licensed and Apache-2.0-licensed sources; the license tags above describe the respective source material, not a replacement license for every row.