yashmarathe/eka-rl

---
license: apache-2.0
language:
- en
tags:
- math
- reasoning
- reinforcement-learning
- grpo
- eka
size_categories:
- 100K<n<1M
dataset_info:
features:
- name: problem
dtype: string
- name: answer
dtype: string
- name: domain
dtype: string
- name: solve_rate
dtype: float64
splits:
- name: train
num_bytes: 74008065
num_examples: 251122
download_size: 44604381
dataset_size: 74008065
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

# eka-rl

A dataset of **251,122 verified math problems** designed for reinforcement learning training of math-reasoning language models. Each problem has a verified correct answer, enabling straightforward binary outcome rewards (correct / wrong) without a process reward model or verifier LLM.

Used with **[eka-finetune](https://github.com/yash-marathe/eka-mono/tree/main/packages/eka-finetune)** to train models via GRPO (Group Relative Policy Optimisation).

---

## Dataset Summary

| Property | Value |
|---|---|
| Problems | 251,122 |
| Split | `train` only |
| Answer format | Plain text / LaTeX |
| Domains | Algebra, Geometry, Number Theory, Combinatorics, Calculus, and more |
| Difficulty range | `solve_rate` 0.0 (hardest) → 1.0 (easiest) |

---

## Columns

| Column | Type | Description |
|---|---|---|
| `problem` | `string` | Math problem statement in LaTeX |
| `answer` | `string` | Verified ground-truth answer |
| `domain` | `string` | Math domain tags (e.g. `['Algebra -> Equations']`) |
| `solve_rate` | `float64` | Fraction of model attempts that are correct — proxy for problem difficulty |

---

## Usage

### Load the full dataset

```python
from datasets import load_dataset

ds = load_dataset("yashmarathe/eka-rl", split="train")
print(ds[0])
# {
# "problem": "Find all integers n such that n^2 + 3 is divisible by 7.",
# "answer": "n ≡ 2 or 5 (mod 7)",
# "domain": "['Mathematics -> Number Theory -> Congruences']",
# "solve_rate": 0.25
# }
```

### Filter by difficulty

`solve_rate` is a useful curriculum signal — start with easier problems and gradually increase difficulty:

```python
from datasets import load_dataset

ds = load_dataset("yashmarathe/eka-rl", split="train")

# Medium difficulty (20–60% solve rate)
medium = ds.filter(lambda x: 0.2 <= float(x["solve_rate"]) <= 0.6)

# Hard only (< 30% solve rate)
hard = ds.filter(lambda x: float(x["solve_rate"]) < 0.3)

print(f"Medium: {len(medium)} problems")
print(f"Hard: {len(hard)} problems")
```

### Use with eka-finetune GRPO training

This dataset is the default for `train_grpo.py` in [eka-finetune](https://github.com/yash-marathe/eka-mono/tree/main/packages/eka-finetune):

```bash
# configs/rl_config.yaml
dataset_name: "yashmarathe/eka-rl"
num_samples: 50000
min_solve_rate: 0.0
max_solve_rate: 0.8 # exclude trivially easy problems

# Run GRPO training
python3 train_grpo.py --config configs/rl_config.yaml
```

---

## Difficulty Distribution

The `solve_rate` column reflects empirical difficulty — problems with a low solve rate are harder:

| Difficulty | Solve rate range | Approx. problems |
|---|---|---|
| Very Hard | 0.0 – 0.1 | ~40K |
| Hard | 0.1 – 0.3 | ~70K |
| Medium | 0.3 – 0.6 | ~80K |
| Easy | 0.6 – 0.8 | ~40K |
| Trivial | 0.8 – 1.0 | ~20K |

For RL training, filtering to `max_solve_rate=0.8` removes trivially easy problems that provide no learning signal (the model already solves them correctly, so the reward is constant across all rollouts).

---

## Answer Verification

Answers can be verified symbolically using `sympy`:

```python
import sympy
from sympy.parsing.latex import parse_latex

def verify(predicted: str, ground_truth: str) -> bool:
try:
diff = sympy.simplify(parse_latex(predicted) - parse_latex(ground_truth))
return diff == 0
except Exception:
# Fall back to normalised string match
import re
norm = lambda s: re.sub(r"\s+", "", s).lower()
return norm(predicted) == norm(ground_truth)
```

This is exactly the verification logic used in `train_grpo.py`'s correctness reward.

---

## Domains

Problems span the full range of competition and undergraduate mathematics:

- **Algebra** — equations, inequalities, polynomials, sequences
- **Geometry** — Euclidean, coordinate, trigonometry
- **Number Theory** — divisibility, congruences, primes
- **Combinatorics** — counting, probability, graph theory
- **Calculus** — limits, derivatives, integrals, series
- **Linear Algebra** — matrices, eigenvalues, vector spaces

---

## Related

- **Model:** [yashmarathe/Eka-4B](https://huggingface.co/yashmarathe/Eka-4B) — 4B reasoning model trained with this dataset
- **Training code:** [eka-finetune](https://github.com/yash-marathe/eka-mono/tree/main/packages/eka-finetune) — GRPO training script (`train_grpo.py`)
- **SFT dataset:** [yashmarathe/OpenMathReasoning](https://huggingface.co/datasets/yashmarathe/OpenMathReasoning) — 1M CoT samples for supervised fine-tuning

---

## License

Apache 2.0