# Model Card for CodeFuse-13B-4K

![Creation Approach](LOGO.png)


[[中文]](#chinese)    [[English]](#english)



<a id="english"></a>


## Model Description
CodeFuse-13B is a 13 billion parameter code generation model trained on the GPT-NeoX framework, capable of handling code sequences of up to 4096 characters. This model was pretrained on a dataset consisting of 1000B token code, Chinese, and English data, covering over 40 programming languages. To further enhance the effectiveness and quality of the generated code, the model was fine-tuned on the CodeFuse-Evol-instruction-66k dataset, enabling it to produce more accurate, efficient, and compliant code. Pass@1 achieved 37.1% on the HumanEval evaluation set(BeamSearch strategy, BeamSize=3).


## Requirements
* Python 3.8 or above.
* PyTorch 1.12 or above, with a recommendation for 2.0 or above.
* Transformers 4.24.0 or above.
* It is advisable to use CUDA 11.4 or above (for GPU users and flash-attention users, this option should be considered).


## Quickstart

```
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("CodeFuse-13B")
model = AutoModelForCausalLM.from_pretrained("CodeFuse-13B", torch_dtype="auto", device_map="auto")

input_ids = encode("def quick_sort(array):\n", return_tensors="pt")
output_ids = model.generate(input_ids, max_new_tokens=200, num_beams=3, num_return_sequences=1, repetition_penalty=1.2)

print(tokenizer.decode(output_idss[0]))
```

<a id="chinese"></a>

## 简介
CodeFuse-13B是基于GPT-NeoX框架训练的13B参数代码生成模型，能够处理4096个字符的代码序列。该模型在1000B Token的代码、中文、英文数据数据集上进行预训练，覆盖超过40种编程语言。为了进一步提升生成代码的效果和质量，该模型还在CodeFuse-Evol-instruction-66k数据集上进行了微调，使得该模型能够生成更加准确、高效、符合要求的代码。在HumanEval评测集上Pass@1达到37.1%(采用BeamSearch解码，其中BeamSize=3)。


## 要求
* python 3.8及以上版本
* pytorch 1.12及以上版本，推荐2.0及以上版本
* transformers 4.24.0及以上版本
* 建议使用CUDA 11.4及以上（GPU用户、flash-attention用户等需考虑此选


## 快速使用

```
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("CodeFuse-13B")
model = AutoModelForCausalLM.from_pretrained("CodeFuse-13B", torch_dtype="auto", device_map="auto")

input_ids = encode("def quick_sort(array):\n", return_tensors="pt")
output_ids = model.generate(input_ids, max_new_tokens=200, num_beams=3, num_return_sequences=1, repetition_penalty=1.2)

print(tokenizer.decode(output_idss[0]))
```