DeepCoder 14B This LOCAL Opensource AI Coding MODEL is CRAZY!



AI Summary

Overview

  • Introduction to Deepcoder 14B, a coding-focused LLM.
  • Specifically designed for code generation and understanding with 14 billion parameters.

Key Features

  • Achieves 60.6% pass accuracy on Live Codebench, an 8% improvement over its base model.
  • Outperforms OpenAI’s 03 mini model while having fewer parameters.
  • Open weights available on Olma for easy accessibility.

Fine-Tuning Process

  • Fine-tuned from Deepseek R1 distilled model using Reinforcement Learning.
  • Utilized 24,000 unique problem test pairs from multiple sources.
  • Training employed Generalized Reward Policy Optimization and Direct Preference Optimization techniques.

Performance Improvements

  • Enhanced capabilities for handling long context outputs.
  • Generalizes well at 64,000 token context length, maintaining a 60.6% accuracy.
  • Use of overlong filtering during training aids performance on longer contexts.

Training Verification

  • Employs code sandbox execution for reinforcement learning evaluation.
  • Tested across 1024 different problems with multiple unit tests.

Accessibility

  • Available on Olma; easy to set up and run locally.
  • Users can select between the 1.5B and 14B parameter models.

User Experience

  • Initial tests show improved performance over the previous distilled variant for complex tasks.
  • However, responses can be slow, especially for simpler tasks due to deep reasoning processes.
  • Weakness observed in tool calling/friction with interfaces but works well in conversational settings.