Member-only story
DeepCoder-14B-Preview:
Open-Source Coding AI That Rivals Proprietary Models
Together, AI and Agentica have released the DeepCoder-14B-Preview, a new open-source model capable of coding performance at the same level as the very best commercial systems. It’s a breakthrough development in the use of reinforcement learning (RL) to create code, an arena that’s typically fallen behind efforts in such domains as mathematical reasoning.
Solving the Data Problem in RL for Code
Training strong coding models with RL has long been hindered by one central issue: good-quality, verifiable datasets are rare. Available resources — such as APPS, TACO, CodeContests, KodCode, and LeetCode — are too basic, noisy, or insufficient in unit tests to be a good RL benchmark.
To address this, the DeepCoder team created a new training set of 24,000 validated problems by combining hand-curated problems from TACO and PrimeIntellect’s SYNTHETIC-1 with a subset of LiveCodeBench problems. A custom pipeline of filters made sure that each problem:
- Had a verifiable reference solution,
- Passed all unit tests,
- Included at least five unit tests (to prevent “reward hacking”), and
- Was deduplicated to avoid contamination with test…
