DeepSeek-Coder - An open-source code language model series for code generation and completion.

## Definition of DeepSeek-Coder DeepSeek-Coder is an open-source series of code language models designed for code generation and completion. It is trained on 2 trillion tokens, with 87% being code and 13% natural language (English and Chinese). The models range from 1B to 33B parameters and support 67 programming languages. ## Supported Programming Languages DeepSeek-Coder supports 67 programming languages, including Python, Java, C++, JavaScript, Ruby, Rust, and many others. A full list is available in the project documentation. ## Definition of DeepSeek-Coder - **Training Data**: 2 trillion tokens (87% code, 13% natural language). - **Model Sizes**: 1B, 5.7B, 6.7B, and 33B parameters. - **Performance**: State-of-the-art results on benchmarks like HumanEval, MultiPL-E, MBPP, DS-1000, and APPS. - **Functionality**: Code completion, insertion, chat model inference, and repository-level code generation. - **Window Size**: 16K tokens for project-level context. ## Code Completion with DeepSeek-Coder DeepSeek-Coder provides code completion through its base models (e.g., `deepseek-ai/deepseek-coder-6.7b-base`). Users can set `eos_token_id` to 32014 for instruct-based completion. The model supports project-level context with a 16K token window. ## Fine-Tuning in DeepSeek-Coder Yes, DeepSeek-Coder supports fine-tuning. The script `finetune_deepseekcoder.py` is provided, compatible with DeepSpeed. Users can fine-tune on datasets like `Evol-Instruct-Code-80k-v1` for specialized tasks. ## Official GitHub Repository The official GitHub repository for DeepSeek-Coder is [https://github.com/deepseek-ai/DeepSeek-Coder](https://github.com/deepseek-ai/DeepSeek-Coder). It includes documentation, demos, and usage instructions. ## DeepSeek-Coder and RPA While DeepSeek-Coder itself is a code-generation model, it can be integrated into broader AI agent frameworks for RPA applications. It does not directly include RPA-specific features like browser automation but can generate code for such workflows when combined with other tools. ## Running a Local Demo Users can run a local demo by executing `app.py` from the `demo` folder in the repository. Dependencies must first be installed via `pip install -r requirements.txt`. An online demo is also available on Hugging Face Spaces. ## Inference Methods DeepSeek-Coder supports inference via: - **vLLM**: For efficient text and chat completion. - **GGUF Quantization**: Compatible with `llama.cpp`. - **GPTQ**: Supported via `exllamav2` with HuggingFace Tokenizer integration. - **Direct API Calls**: Using the provided model endpoints. ### Citation sources: - [DeepSeek-Coder](https://github.com/deepseek-ai/DeepSeek-Coder) - Official URL Updated: 2025-04-01