TeamClaude/GRPO-Fine-Tuned

GSM8K GRPO Dataset for Qwen3-4B Post-Training

A clean, GRPO-ready dataset derived from GSM8K for post-training the Qwen3-4B base model using Group Relative Policy Optimization (GRPO).

Dataset Purpose

This dataset is designed for the GRPO stage of post-training, where the model learns to produce correct mathematical reasoning through reward-based optimization. The key design principles are:

1. Verifiable answers: Every example has a single, unambiguous numeric answer
2. Clean reward extraction: The #### <answer> format makes automated answer checking trivial
3. Moderate difficulty: Balanced to provide useful learning signal (not too easy, not too hard)
4. Format consistency: Standardized prompt templates teach the model the exact output format

Dataset Splits

Split	Size	Description
v1_clean	7,473	All GSM8K train examples with clean numeric answers (full baseline)
v2_medium	6,100	Difficulty-balanced: 20% easy / 65% medium / 15% hard
v3_augmented	18,300	v2 examples x 3 prompt template variants (format robustness)

Recommended for GRPO training: Use v2_medium as the primary training set.

Column Descriptions

Column	Type	Description
prompt	string	The formatted prompt including question and answer-format instruction
question	string	The original GSM8K question text
gold_solution	string	The original GSM8K step-by-step solution (for debugging, not for training)
gold_answer	string	The correct final answer as a string (e.g., "72")
gold_answer_float	float	The correct final answer as a float (for reward computation)
difficulty	string	Difficulty bucket: "easy", "medium", or "hard"
num_numbers_in_question	int	Count of numeric values in the question
num_solution_steps	int	Number of reasoning steps in the gold solution
prompt_template	string	Which prompt template was used: "standard", "variant_a", or "variant_b"
source	string	Always "gsm8k"

Usage with TRL GRPOTrainer

from datasets import load_dataset
from trl import GRPOTrainer, GRPOConfig
import re

# Load the recommended split
dataset = load_dataset("TeamClaude/GRPO-Fine-Tuned", split="v2_medium")

# Define reward function
def math_reward_func(completions, gold_answer, **kwargs):
    """Check if model's final answer matches the gold answer."""
    rewards = []
    for completion, answer in zip(completions, gold_answer):
        # Extract the model's answer from #### format
        match = re.search(r"####\s*(.+)", completion)
        if match:
            model_answer = match.group(1).strip()
            model_answer = model_answer.replace(",", "").replace("$", "")
            try:
                model_val = float(model_answer)
                gold_val = float(answer)
                rewards.append(1.0 if abs(model_val - gold_val) < 0.01 else 0.0)
            except ValueError:
                rewards.append(0.0)
        else:
            rewards.append(0.0)
    return rewards

# Optional: format reward to encourage #### output
def format_reward_func(completions, **kwargs):
    """Reward completions that contain the #### format."""
    return [0.5 if "####" in c else 0.0 for c in completions]

# Train
trainer = GRPOTrainer(
    model="your-sft-checkpoint",
    reward_funcs=[math_reward_func, format_reward_func],
    train_dataset=dataset,
    args=GRPOConfig(
        per_device_train_batch_size=4,
        num_generations=8,
        max_completion_length=512,
    ),
)
trainer.train()

Filtering Methodology

Answer Filtering

• Extracted final answer from GSM8K's #### <answer> format
• Normalized: removed $, ,, % symbols; converted fractions to decimals
• Rejected: answers with units, dates, times, ratios, mixed numbers, or non-numeric text
• Result: 100% of GSM8K train passed (GSM8K is already very clean)

Difficulty Classification

Based on three features:

• Easy: <=2 solution steps, <=3 numbers in question, <=40 word question
• Hard: >=6 solution steps, >=7 numbers, or >=120 word question
• Medium: Everything else

Difficulty Rebalancing (v2_medium)

Target distribution: 20% easy / 65% medium / 15% hard. Medium examples subsampled to match proportions. Easy and hard examples kept in full.

Format Augmentation (v3_augmented)

Each v2 example is replicated with 3 prompt instruction variants:

• standard: "Solve the problem step by step..."
• variant_a: "Work through the arithmetic carefully..."
• variant_b: "Answer the question step by step..."

Source

Built from the train split of openai/gsm8k (7,473 examples).

License

This dataset inherits the MIT license from GSM8K.