sungyub/code-contests-plus-verl

Code Contests Plus (VERL Format)

This dataset contains 8,432 competitive programming problems from the Code-Contests-Plus dataset, converted to VERL format for reinforcement learning applications. Each problem includes test cases validated through sandbox execution.

Source: ByteDance-Seed/Code-Contests-Plus (1x config)

License: MIT

Dataset Structure

The dataset follows the VERL format with the following fields:

• —data_source (string): Dataset source identifier ("code-contests-plus")
• —prompt (list): Chat template format with role/content structure containing the coding problem
• —ability (string): Task category ("code")
• —reward_model (dict): Evaluation information
• —style: Evaluation method ("rule")
• —ground_truth: JSON-encoded test cases with input/output pairs
• —extra_info (dict): Additional metadata
• —index: Example index from original dataset

Test Case Format

Each problem includes test cases in the reward_model.ground_truth field, stored as JSON with the following structure:

{
  "inputs": ["3\n1 2 3\n"],
  "outputs": ["6\n"]
}

The format consists of two parallel arrays:

• —inputs: Array of input strings for each test case
• —outputs: Array of expected output strings corresponding to each input

Each problem typically contains between 1 and 32 test cases, validated through sandbox execution during dataset creation.

Data Processing

The dataset was created through a multi-step processing pipeline:

1. Test Case Extraction

• —Extracted public test cases from the original dataset
• —Validated format and executability
• —Filtered problems without valid test cases

2. Sandbox Validation

• —Each problem's test cases were validated using a sandbox environment
• —Test input/output pairs verified for correctness
• —Only problems with passing validation were included

3. Size Filtering

• —Applied 10MB size limit to test case JSON (encoded)
• —Removed overly large problems to ensure efficient processing
• —Balanced dataset quality and usability

Processing Statistics

• —Total input examples: 11,690
• —Successfully processed: 8,432 (72.1% success rate)
• —Total filtered: 3,258 (27.9%)
• —No test cases: 54 (0.5%)
• —Size filtered (>10MB): 3,204 (27.4%)
• —Processing time: 69 minutes
• —Configuration used: 1x (standard difficulty)

Usage

from datasets import load_dataset
import json

# Load the dataset
dataset = load_dataset("sungyub/code-contests-plus-verl")

# Access an example
example = dataset['train'][0]

# Get the problem description
problem = example['prompt'][0]['content']
print("Problem:", problem)

# Parse test cases
ground_truth = json.loads(example['reward_model']['ground_truth'])
inputs = ground_truth['inputs']
outputs = ground_truth['outputs']

print(f"\nNumber of test cases: {len(inputs)}")
print(f"First input: {repr(inputs[0])}")
print(f"Expected output: {repr(outputs[0])}")

Example Problem

Problem Description:

Twins

square1001 and E869120 are twins, but they are not identical twins...

Test Case:

Input:  ""
Output: "square1001"

Statistics

• —Total examples: 8,432
• —Average test cases per problem: ~10-15
• —Test case range: 1-32 per problem
• —Dataset size: ~10 GB uncompressed, ~10 GB compressed (includes test cases)
• —Format: Parquet (11 shards, ~1GB each)
• —Schema: VERL-compatible

Data Quality

All problems in this dataset have been validated to ensure:

1. 1.Valid test cases: Each problem has at least one valid test case
2. 2.Correct input/output pairs: Test cases verified through sandbox execution
3. 3.Size constraints: Test cases are within reasonable size limits (≤10MB)
4. 4.Format consistency: All examples follow the same schema structure

Conversion Script

The dataset was created using preprocess_codecontests_verl.py:

# Standard conversion (used for this dataset)
python preprocess_codecontests_verl.py \
  --dataset-id ByteDance-Seed/Code-Contests-Plus \
  --config 1x \
  --output-dir ./codecontests_verl_full \
  --sandbox-url http://localhost:8080/run_code \
  --batch-size 100

# Process with different configuration
python preprocess_codecontests_verl.py \
  --dataset-id ByteDance-Seed/Code-Contests-Plus \
  --config 2x \
  --output-dir ./codecontests_verl_2x \
  --sandbox-url http://localhost:8080/run_code \
  --batch-size 100

# Process limited samples for testing
python preprocess_codecontests_verl.py \
  --dataset-id ByteDance-Seed/Code-Contests-Plus \
  --config 1x \
  --output-dir ./codecontests_test \
  --sandbox-url http://localhost:8080/run_code \
  --max-examples 100

Related Datasets

• —Code Contests Plus (Original): Original dataset with competitive programming problems
• —Skywork-OR1-Code-VERL: Similar VERL-format dataset with 14,057 coding problems

Additional Information

For more information about VERL format and usage in reinforcement learning, see:

• —VERL Documentation
• —VERL GitHub Repository

Citation

If you use this dataset, please cite the original Code-Contests-Plus dataset:

@misc{code-contests-plus,
  title={Code-Contests-Plus},
  author={ByteDance-Seed},
  year={2024},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/ByteDance-Seed/Code-Contests-Plus}
}

License

This dataset is released under the MIT License, following the license of the original Code-Contests-Plus dataset.