DeepSeek Coder: Let the Code Write Itself
[ Homepage] | [🤖 Chat with DeepSeek Coder] | [🤗 Models Download] | [Discord] | [WeChat (微信)]
- **Massive Training Data**: Trained from scratch on 2T tokens, including 87% code and 13% linguistic data in both English and Chinese languages. - **Highly Flexible & Scalable**: Offered in model sizes of 1B, 5.7B, 6.7B and 33B, enabling users to choose the setup most suitable for their requirements. - **Superior Model Performance**: State-of-the-art performance among publicly available code models on HumanEval, MultiPL-E, MBPP, DS-1000, and APPS benchmarks. - **Advanced Code Completion Capabilities**: A window size of 16K and a fill-in-the-blank task, supporting project-level code completion and infilling tasks. #### Supported Programming Languages `['ada', 'agda', 'alloy', 'antlr', 'applescript', 'assembly', 'augeas', 'awk', 'batchfile', 'bluespec', 'c', 'c-sharp', 'clojure', 'cmake', 'coffeescript', 'common-lisp', 'cpp', 'css', 'cuda', 'dart', 'dockerfile', 'elixir', 'elm', 'emacs-lisp', 'erlang', 'f-sharp', 'fortran', 'glsl', 'go', 'groovy', 'haskell', 'html', 'idris', 'isabelle', 'java', 'java-server-pages', 'javascript', 'json', 'julia', 'jupyter-notebook', 'kotlin', 'lean', 'literate-agda', 'literate-coffeescript', 'literate-haskell', 'lua', 'makefile', 'maple', 'markdown', 'mathematica', 'matlab', 'ocaml', 'pascal', 'perl', 'php', 'powershell', 'prolog', 'protocol-buffer', 'python', 'r', 'racket', 'restructuredtext', 'rmarkdown', 'ruby', 'rust', 'sas', 'scala', 'scheme', 'shell', 'smalltalk', 'solidity', 'sparql', 'sql', 'stan', 'standard-ml', 'stata', 'systemverilog', 'tcl', 'tcsh', 'tex', 'thrift', 'typescript', 'verilog', 'vhdl', 'visual-basic', 'xslt', 'yacc', 'yaml', 'zig']` ### 2. Evaluation Results We evaluate DeepSeek Coder on various coding-related benchmarks. Only `pass@1` results on HumanEval (Python and Multilingual), MBPP, and DS-1000 are reported here:
The result shows that DeepSeek-Coder-Base-33B significantly outperforms existing open-source code LLMs. Compared with CodeLlama-34B, it leads by 7.9%, 9.3%, 10.8% and 5.9% respectively on HumanEval Python, HumanEval Multilingual, MBPP and DS-1000. Surprisingly, our DeepSeek-Coder-Base-7B reaches the performance of CodeLlama-34B. The DeepSeek-Coder-Instruct-33B model after instruction tuning outperforms GPT35-turbo on HumanEval and achieves comparable results with GPT35-turbo on MBPP. More evaluation details can be found in the [Detailed Evaluation](#6-detailed-evaluation-results). ### 3. Procedure of Data Creation and Model Training #### Data Creation - Step 1: Collect code data from GitHub and apply the same filtering rules as [StarCoder Data](https://github.com/bigcode-project/bigcode-dataset) to filter data. - Step 2: Parsing the dependencies of files within the same repository to rearrange the file positions based on their dependencies. - Step 3: Concatenating dependent files to form a single example and employ repo-level minhash for deduplication. - Step 4: Further filtering out low-quality code, such as codes with syntax errors or poor readability. #### Model Training - Step 1: Initially pre-trained with a dataset consisting of 87% code, 10% code-related language (Github Markdown and StackExchange), and 3% non-code-related Chinese language. Models are pre-trained using 1.8T tokens and a 4K window size in this step. - Step 2: Further Pre-training using an extended 16K window size on an additional 200B tokens, resulting in foundational models (**DeepSeek-Coder-Base**). - Step 3: Instruction Fine-tuning on 2B tokens of instruction data, resulting in instruction-tuned models (**DeepSeek-Coder-Instruct**). ### 4. How to Use Before proceeding, you'll need to install the necessary dependencies. You can do this by running the following command: ``` pip install -r requirements.txt ``` A demo is also available on the [🤗 Hugging Face Space](https://huggingface.co/spaces/deepseek-ai/deepseek-coder-33b-instruct), and you can run the demo locally using `app.py` in the [demo](https://github.com/deepseek-ai/deepseek-coder/tree/main/demo) folder. (Thanks to all the HF team for their support) Here are some examples of how to use our model. #### 1) Code Completion ```python from transformers import AutoTokenizer, AutoModelForCausalLM import torch tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/deepseek-coder-6.7b-base", trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained("deepseek-ai/deepseek-coder-6.7b-base", trust_remote_code=True, torch_dtype=torch.bfloat16).cuda() input_text = "#write a quick sort algorithm" inputs = tokenizer(input_text, return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_length=128) print(tokenizer.decode(outputs[0], skip_special_tokens=True)) ``` This code will output the following result: ``` def quick_sort(arr): if len(arr) <= 1: return arr pivot = arr[0] left = [] right = [] for i in range(1, len(arr)): if arr[i] < pivot: left.append(arr[i]) else: right.append(arr[i]) return quick_sort(left) + [pivot] + quick_sort(right) ``` #### 2) Code Insertion ``` … ``` This code will output the following result: ``` for i in range(1, len(arr)): ``` #### 3) Chat Model Inference ``` … ``` This code will output the following result: ``` … ``` If you don't want to use the provided API `apply_chat_template` which loads the template from `tokenizer_config.json`, you can use the following template to chat with our model. Replace the `['content']` with your instructions and the model's previous (if any) responses, then the model will generate the response to the currently given instruction. ``` You are an AI programming assistant, utilizing the DeepSeek Coder model, developed by DeepSeek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer. ### Instruction: ['content'] ### Response: ['content'] <|EOT|> ### Instruction: ['content'] ### Response: ``` #### 4) Repository Level Code Completion ``` … ``` --- In the following scenario, the DeepSeek-Coder-6.7B model effectively calls a class **IrisClassifier** and its member function from the `model.py` file, and also utilizes functions from the `utils.py` file, to correctly complete the **main** function in the `main.py` file for model training and evaluation. ### 5. How to Fine-tune DeepSeek-Coder We provide script `finetune/finetune_deepseekcoder.py` for users to finetune our models on downstream tasks. The script supports the training with [DeepSpeed](https://github.com/microsoft/DeepSpeed). You need install required packages by: ```bash pip install -r finetune/requirements.txt ``` Please follow [Sample Dataset Format](https://huggingface.co/datasets/nickrosh/Evol-Instruct-Code-80k-v1) to prepare your training data. Each line is a json-serialized string with two required fields `instruction` and `output`. After data preparation, you can use the sample shell script to finetune `deepseek-ai/deepseek-coder-6.7b-instruct`. Remember to specify `DATA_PATH`, `OUTPUT_PATH`. And please choose appropriate hyper-parameters(e.g., `learning_rate`, `per_device_train_batch_size`) according to your scenario. ``` … ``` ### 6. Detailed Evaluation Results The reproducible code for the following evaluation results can be found in the [Evaluation](https://github.com/deepseek-ai/deepseek-coder/tree/main/Evaluation) directory. #### 1) Multilingual HumanEval Benchmark #### 2) MBPP Benchmark #### 3) DS-1000 Benchmark #### 4) Program-Aid Math Reasoning Benchmark ### Inference with vLLM You can also employ [vLLM](https://github.com/vllm-project/vllm) for high-throughput inference. **Text Completion** ``` … ``` **Chat Completion** ``` … ``` ### 7. Q&A #### Could You Provide the tokenizer.model File for Model Quantization? DeepSeek Coder utilizes the [HuggingFace Tokenizer](https://huggingface.co/docs/tokenizers/index) to implement the Bytelevel-BPE algorithm, with specially designed pre-tokenizers to ensure optimal performance. Currently, there is no direct way to convert the tokenizer into a SentencePiece tokenizer. We are contributing to the open-source quantization methods facilitate the usage of HuggingFace Tokenizer. ##### GGUF(llama.cpp) We have submitted a [PR](https://github.com/ggerganov/llama.cpp/pull/4070) to the popular quantization repository [llama.cpp](https://github.com/ggerganov/llama.cpp) to fully support all HuggingFace pre-tokenizers, including ours. While waiting for the PR to be merged, you can generate your GGUF model using the following steps: ```bash git clone https://github.com/DOGEwbx/llama.cpp.git cd llama.cpp git checkout regex_gpt2_preprocess # set up the environment according to README make python3 -m pip install -r requirements.txt # generate GGUF model python convert-hf-to-gguf.py --outfile --model-name deepseekcoder # use q4_0 quantization as an example ./quantize q4_0 ./main -m -n 128 -p ``` ##### GPTQ(exllamav2) `UPDATE:`[exllamav2](https://github.com/turboderp/exllamav2) has been able to support Huggingface Tokenizer. Please pull the latest version and try out. Remember to set RoPE scaling to 4 for correct output, more discussion could be found in this [PR](https://github.com/turboderp/exllamav2/pull/189). #### How to use the deepseek-coder-instruct to complete the code? Although the deepseek-coder-instruct models are not specifically trained for code completion tasks during supervised fine-tuning (SFT), they retain the capability to perform code completion effectively. To enable this functionality, you simply need to adjust the eos_token_id parameter. Set the eos_token_id to 32014, as opposed to its default value of 32021 in the deepseek-coder-instruct configuration. This modification prompts the model to recognize the end of a sequence differently, thereby facilitating code completion tasks. ### 8. Resources [awesome-deepseek-coder](https://github.com/deepseek-ai/awesome-deepseek-coder) is a curated list of open-source projects related to DeepSeek Coder. ### 9. License This code repository is licensed under the MIT License. The use of DeepSeek Coder models is subject to the Model License. DeepSeek Coder supports commercial use. See the [LICENSE-CODE](LICENSE-CODE) and [LICENSE-MODEL](LICENSE-MODEL) for more details. ### 10. Citation ``` @misc{deepseek-coder, author = {Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y.K. Li, Fuli Luo, Yingfei Xiong, Wenfeng Liang}, title = {DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence}, journal = {CoRR}, volume = {abs/2401.