What is DeepCoder-14B-Preview
DeepCoder-14B-Preview is a large code generation model jointly open-sourced by Agentica and Together AI, fine-tuned based on Deepseek-R1-Distilled-Qwen-14B. DeepCoder-14B-Preview is trained with distributed reinforcement learning (RL), performing excellently on code generation tasks, particularly achieving 60.6% accuracy on LiveCodeBench, comparable to OpenAI's o3-mini . The model open-sources the training dataset, code, training logs, and system optimizations, promoting the application of reinforcement learning (RL) in large language models (LLM), lowering the barrier to RL training, and fostering community development.
Main Features of DeepCoder-14B-Preview
- High-Quality Code Generation: Generates high-quality, runnable code, suitable for multiple programming languages and scenarios.
- Code Problem Solving: Solves complex programming problems, including algorithm design, data structure optimization, etc.
- Code Completion and Optimization: Provides code completion functionality, helping developers quickly complete code writing and optimize existing code to improve efficiency.
- Unit Test Generation: Automatically generates unit test code, ensuring the accuracy and reliability of generated code.
- Code Debugging Assistance: Helps developers locate and fix errors in code, improving development efficiency.
- Cross-Platform Applicability: Supports multiple programming environments and platforms, with broad applicability.
Technical Principles of DeepCoder-14B-Preview
- Base Model: Based on Deepseek-R1-Distilled-Qwen-14B, a distilled and optimized 14-billion-parameter pre-trained model, with powerful language understanding and generation capabilities.
- Reinforcement Learning Fine-Tuning: Fine-tunes the base model based on distributed reinforcement learning (RL). Reinforcement learning guides the model to generate higher-quality code based on a reward mechanism, ensuring the accuracy and efficiency of the code.
- High-Quality Dataset: Trained with 24K rigorously filtered verifiable programming problems, with data sources including TACO Verified, PrimeIntellect's SYNTHETIC-1 dataset, and problems submitted to LiveCodeBench.
- Reward Function Design: Based on a sparse outcome reward model (ORM), rewards are given only when the generated code passes all sampled unit tests, preventing the model from obtaining rewards by memorizing test cases.
- Context Extension Technology: Iterative context extension technology, where the model starts learning from shorter context lengths and gradually generalizes to longer contexts, ultimately achieving 60.6% accuracy in 64K context.
- System Optimization: Introduces verl-pipeline, based on pipeline technology to accelerate the training process, reduce training time, and improve training efficiency.
Project Address of DeepCoder-14B-Preview
- Project Official Website：https://pretty-radio-b75.notion.site/DeepCoder-A-Fully-Open-Source-14B-Coder
- HuggingFace Model Repository：https://huggingface.co/agentica-org/DeepCoder-14B-Preview
Application Scenarios of DeepCoder-14B-Preview
- Code Generation and Automated Programming: Quickly generates high-quality code, reducing the time and workload of manual code writing, and improving development efficiency. Suitable for various programming languages and frameworks, helping developers quickly start projects.
- Algorithm Competitions and Problem Solving: In algorithm competitions (such as Codeforces), help participants quickly understand problems and generate efficient solutions to improve competition performance.
- Code Optimization and Refactoring: Optimize and refactor existing code to improve readability, performance, and maintainability. Help developers identify and fix potential code issues.
- Education and Learning Assistance: As a programming education tool, help students understand and practice programming concepts, provide code examples and solutions, and assist in learning programming languages and algorithms.
- Software Development and Testing: Generate unit test code to ensure software quality; assist with code debugging during development, help developers quickly locate and resolve issues, and improve the overall efficiency of software development.