AB
AB
AiBoss

All Categories

project

DeepCoder-14B-Preview - A code generation model jointly developed by Agentica and Together AI.

DeepCoder-14B-Preview is a large-scale code generation model jointly open-sourced by Agentica and Together AI, finely tuned from Deepseek-R1-Distilled-Qwen-14B. DeepCoder-14B-Preview is based on...

What is DeepCoder-14B-Preview?

DeepCoder-14B-Preview is a large-scale code generation model jointly open-sourced by Agentica and Together AI, finely tuned from Deepseek-R1-Distilled-Qwen-14B. Trained using distributed reinforcement learning (RL), DeepCoder-14B-Preview performs exceptionally well on code generation tasks, achieving a 60.6% accuracy on LiveCodeBench, comparable to OpenAI's o3-mini. The model's training dataset, code, training logs, and system optimizations have been open-sourced, promoting the application of reinforcement learning (RL) in large-scale language models (LLMs), lowering the barrier to RL training, and fostering community development.

Main functions of DeepCoder-14B-Preview

  • High-quality code generationGenerates high-quality, runnable code suitable for various programming languages and scenarios.
  • Code problem solvedSolving complex programming problems, including algorithm design and data structure optimization.
  • Code completion and optimizationIt provides code completion functionality to help developers quickly complete code writing and optimize existing code to improve efficiency.
  • Unit test generationAutomatically generates unit test code, ensuring the accuracy and reliability of the generated code.
  • Code debugging aidIt helps developers locate and fix errors in their code, improving development efficiency.
  • Cross-platform compatibilityIt supports multiple programming environments and platforms, making it widely applicable.

The technical principles of DeepCoder-14B-Preview

  • Basic ModelBased on Deepseek-R1-Distilled-Qwen-14B, a pre-trained model with 14 billion parameters optimized by distillation, it has powerful language understanding and generation capabilities.
  • Reinforcement Learning Fine-tuningThis approach uses distributed reinforcement learning (RL) to fine-tune the base model. Reinforcement learning guides the model to generate higher-quality code based on reward mechanisms, ensuring the accuracy and efficiency of the code.
  • High-quality datasetsThe dataset was trained using 24,000 rigorously selected verifiable programming problems, with data sources including TACO Verified, PrimeIntellect's SYNTHETIC-1 dataset, and problems submitted to LiveCodeBench.
  • Reward function designBased on the sparse result reward model (ORM), a reward is only given when the generated code is tested with all sampled units, avoiding the model from obtaining rewards by memorizing test cases.
  • Context extension technologyThe iterative context expansion technique allows the model to learn from shorter context lengths and gradually generalize to longer contexts, ultimately achieving an accuracy of 60.6% in 64K contexts.
  • System optimizationIntroducing verl-pipeline, which accelerates the training process, reduces training time, and improves training efficiency based on pipeline technology.

The project address for DeepCoder-14B-Preview

Application scenarios of DeepCoder-14B-Preview

  • Code generation and automated programmingIt quickly generates high-quality code, reducing the time and workload of manual coding and improving development efficiency. It is compatible with various programming languages and frameworks, helping developers quickly start projects.
  • Algorithm competitions and problem solvingIn algorithm competitions (such as Codeforces), it helps participants quickly understand the problem and generate efficient solutions, thereby improving their competition performance.
  • Code optimization and refactoringOptimize and refactor existing code to improve readability, performance, and maintainability. Help developers identify and fix potential code problems.
  • Education and learning supportAs a programming education tool, it helps students understand and practice programming concepts, provides code examples and solutions, and assists in learning programming languages and algorithms.
  • Software Development and TestingGenerate unit test code to ensure software quality; assist in code debugging during the development process, help developers quickly locate and solve problems, and improve the overall efficiency of software development.
Back