DeepCoder Preview: 24,000 Coding Problems with Verified Test Cases

24,000 coding problems with verified test cases, released by agentica-org in 2025. The collection aggregates 7.5K TACO problems, 16K synthetic problems from PrimeIntellect, and 600 LiveCodeBench entries for training code generation models.

Use Cases

• Fine-tuning code generation models using the `problem` descriptions and test cases
• Evaluating model performance on recent competitive programming tasks from the LiveCodeBench 2024-2025 split
• Analyzing the impact of synthetic training data from the SYNTHETIC-1 subset on code synthesis accuracy

Strengths

• 24,000 training records with verified test cases
• Includes 7,500 TACO verified problems
• Temporal test split covering August 2024 to February 2025

Limitations

• High proportion of synthetic data (66% of training set)
• Potential for data leakage if models were previously trained on the TACO or Codeforces source datasets

Provenance

Source

TACO, PrimeIntellect (SYNTHETIC-1), LiveCodeBench (v5), and Codeforces (Qwen/CodeElo)

Collection Method

Aggregated from existing verified datasets and synthetic generation

Time Range

2023-05-01 to 2025-02-01

Freshness

Last updated April 2025; includes test problems as recent as February 2025.