microsoft/NextCoder-7B

1---2license: mit3language:4- en5base_model:6- Qwen/Qwen2.5-Coder-7B-Instruct7pipeline_tag: text-generation8library_name: transformers9tags:10- code11- chat12- microsoft13- nextcoder14- selekt15datasets:16- microsoft/NextCoderDataset17- microsoft/NextCoderDataset-Conversational18- bigcode/commitpackft19- bigcode/starcoderdata20---21 22 23# NextCoder-7B24<p align="center">25        <a href="https://github.com/microsoft/NextCoder">GitHub</a>&nbsp&nbsp | &nbsp&nbsp <a href="https://www.microsoft.com/en-us/research/publication/nextcoder-robust-adaptation-of-code-lms-to-diverse-code-edits/">Paper</a> 26</p>27 28> NextCoder: Robust Adaptation of Code LMs to Diverse Code Edits (ICML'2025)29 30## Introduction31 32NextCoder is the latest series of Code-Editing large language models developed using the Qwen2.5-Coder Instruct variants as base and trained with novel Selective Knowledge Transfer finetuning methodology as introduced in the paper. NextCoder family model comes in 3 different sizes 7, 14, 32 billion parameters, to meet the needs of different developers.33Following are the key improvements:34- Significantly improvements in **code editing**, NextCoder-32B has performing on par with GPT-4o on complex benchmarks like Aider-Polyglot with performance increment of 44% from their base model.35- No loss of generalizibility, due to our new finetuning method **SeleKT**36- **Long-context Support** up to 32K tokens.37 38**This repo contains the NextCoder-7B model**, which has the following features:39- Type: Causal Language Models40- Training Stage: Post-training with SeleKT41- Architecture: transformers with RoPE, SwiGLU, RMSNorm, and Attention QKV bias42- Number of Parameters: 7.61B43- Number of Paramaters (Non-Embedding): 6.53B44- Number of Layers: 2845- Number of Attention Heads (GQA): 28 for Q and 4 for KV46  47For more details, please refer to our [blog](), [GitHub](https://github.com/microsoft/NextCoder), [Paper](https://www.microsoft.com/en-us/research/publication/nextcoder-robust-adaptation-of-code-lms-to-diverse-code-edits/).48 49## Requirements50 51The code of NextCoder is based on Qwen2.5 base models which has been in the latest Hugging face `transformers` and we advise you to use the latest version of `transformers`.52 53With `transformers<4.37.0`, you will encounter the following error:54```55KeyError: 'qwen2'56```57 58## Quickstart59 60Here provides a code snippet with `apply_chat_template` to show you how to load the tokenizer and model and how to generate contents.61 62```python63from transformers import AutoModelForCausalLM, AutoTokenizer64 65model_name = "microsoft/NextCoder-7B"66 67model = AutoModelForCausalLM.from_pretrained(68    model_name,69    torch_dtype="auto",70    device_map="auto",71)72tokenizer = AutoTokenizer.from_pretrained(model_name)73 74prompt = """75Fix the following function that divides two numbers to handle all the edge cases:76 77def divide(a, b)78  returm a/b79"""80messages = [81    {"role": "user", "content": prompt}82]83text = tokenizer.apply_chat_template(84    messages,85    tokenize=False,86    add_generation_prompt=True87)88model_inputs = tokenizer([text], return_tensors="pt").to(model.device)89 90generated_ids = model.generate(91    **model_inputs,92    max_new_tokens=102493)94generated_ids = [95    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)96]97 98response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]99```100## Evaluation and Performance101 102| Models | HUMANEVALFIX | CANITEDIT | AIDER | POLYGLOT |103|--------|---------------|-----------|-------|----------|104| QwenCoder-2.5-3B | 73.2 | 37.1 | 36.8 | - |105| QwenCoder-2.5-3B-LoRA | 64.6 | 36.2 | 35.8 | - |106| QwenCoder-2.5-3B-SFT | 76.2 | 32.4 | 30.1 | - |107| **NextCoder-3B** | 75.6 | 42.4 | 37.6 | - |108| QwenCoder-2.5-7B | 73.8 | 48.1 | 59.4 | - |109| QwenCoder-2.5-7B-LoRA | 70.7 | 44.3 | 40.6 | - |110| QwenCoder-2.5-7B-SFT | 70.1 | 36.7 | 48.9 | - |111| **NextCoder-7B** | 81.1 | 50.5 | 65.7 | - |112| QwenCoder-2.5-14B | 87.8 | 58.1 | 66.9 | 9.3 |113| QwenCoder-2.5-14B-LoRA | 78.0 | 50.9 | 66.2 | 5.3 |114| QwenCoder-2.5-14B-SFT | 79.9 | 42.4 | 36.8 | 3.1 |115| **NextCoder-14B** | 89.8 | 60.2 | 72.2 | 12.2 |116| QwenCoder-2.5-32B | **90.2** | 61.0 | 72.9 | 16.4 |117| QwenCoder-2.5-32B-LoRA | 82.3 | 52.4 | 60.2 | 6.7 |118| QwenCoder-2.5-32B-SFT | 81.7 | 49.5 | 66.9 | 8.4 |119| **NextCoder-32B** | 88.9 | **62.4** | **74.7** | **23.6** |120 121*Comparison of base QwenCoder-2.5 models of different sizes and their SELEKT-enhanced versions across three code editing benchmarks.*122 123**Detailed evaluation results are reported in this [📑 paper](https://www.microsoft.com/en-us/research/publication/nextcoder-robust-adaptation-of-code-lms-to-diverse-code-edits/).**124 125## Responsible AI Use126The base models (from the QwenCoder-2.5 family) are suspectible to malicious prompts and may generate or execute harmful code. Our finetuning does not enhance or impede such behaviors. The users should use the models and their outputs responsibly and with caution. Model outputs should be subjected to additional analysis, including manual inspection, and sandboxing before execution.127 128## Citation129 130```bibtex131@inproceedings{aggarwal2025nextcoder,132author = {Aggarwal, Tushar and Singh, Swayam and Awasthi, Abhijeet and Kanade, Aditya and Natarajan, Nagarajan},133title = {NextCoder: Robust Adaptation of Code LMs to Diverse Code Edits},134booktitle = {International Conference on Machine Learning},135year = {2025},136url = {https://www.microsoft.com/en-us/research/publication/nextcoder-robust-adaptation-of-code-lms-to-diverse-code-edits/},137}138```