Nirav-Madhani/meditation-math-seeds

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- math
- meditation
- reinforcement-learning
- GRPO
- seed-data
size_categories:
- 1K<n<10K
---

# Meditation Math Seeds

Training data for the **Meditation** project: self-supervised mathematical introspection as a training phase for language models.

## Dataset Structure

| File | Records | Description |
|------|---------|-------------|
| topics.json | 188 | Math topics across 3 tiers with references |
| seed_meditations.jsonl | 1,504 | Raw meditations generated by Gemini 3.1 Flash Lite |
| seeds_filtered.jsonl | 601 | Quality-filtered meditations (40% pass rate) |
| sft_formatted.jsonl | 601 | Chat-formatted for SFT training |

## Topic Tiers

- **Tier 1** (50%): Fully verifiable with sympy (arithmetic, modular, combinatorics, series, polynomials)
- **Tier 2** (30%): Partially verifiable (calculus, linear algebra, number theory, optimization)
- **Tier 3** (20%): Judge-evaluated (abstract algebra, real analysis, topology, competition math)

## Quality Filters

- Word count: 80-800 words
- Math character density >= 15
- Self-posed problem detected
- Sympy verification: no failed claims (tier 1-2)
- Minimum verified claims: tier 1 >= 2, tier 2 >= 1

## Citation

See the [paper](https://github.com/Nirav-Madhani/MeditationLearning) for details.