Fine-tuning

We only provide the code to fuse instructions. You can use the following GitHub repo for fine-tuning Code Llama: https://github.com/hiyouga/LLaMA-Factory/tree/main

Dataset

download it from huggingface:

• https://huggingface.co/datasets/Pasta009/Instruction-Fusion-Code-v1

Evaluation

Use evalplus to evaluate on python benchmarks: https://github.com/evalplus/evalplus

Use bigcode-evaluation-harness to evalute the performance on MultiPL-E https://github.com/bigcode-project/bigcode-evaluation-harness

Seed Template

{ "dataset": "your dataset name", "id": 0, "instruction": "", }

We use evol-codealpaca-v1 as seeds.

Performance

Method	Size	Open-source	HumanEval	HumanEval+	MBPP	MBPP+
gpt-4-1106-preview	-	-	85.4	81.7	83.0	70.7
gpt-3.5-turbo-1106	-	-	72.6	65.9	81.7	69.4
StarCoder	7B	weight&data	24.4	20.7	33.1	28.8
Mistral	7B	weight	28.7	23.2	50.1	40.9
CodeLlama-Python	7B	weight	37.8	34.1	57.6	45.4
WizardCoder-CLP	7B	weight	48.2	40.9	56.6	47.1
MagicoderS-CLP	7B	weight&data	70.7	66.5	68.4	56.6
CodeLlama-Python	13B	weight	42.7	36.6	61.2	50.9
StarCoder	15B	weight&data	34.1	29.3	55.1	46.1
CodeT5+	16B	weight&data	31.7	26.2	54.6	44.4
CodeGen-Mono	16B	weight&data	32.9	27.4	52.6	43.6
CodeLlama-Python	34B	weight	51.8	42.7	67.2	52.9
WizardCoder-CLP	34B	weight	73.2	64.6	73.2	59.9
IF-CLP	13B	weight&data	73.8	69.5	71.7	61.7
IF-CL	13B	weight&data	74.4	68.3	69.7	59.4
IF-CLP	34B	weight&data	75.6	69.5	73.7	62.7
IF-CL	34B	weight&data	78.7	71.3	71.4	60.7
IF-CL-MC	7B	weight&data	76.2	71.3	70.4	57.9
IF-CL-MC	13B	weight&data	79.3	72.6	69.2	57.4
IF-CL-MC	34B	weight&data	82.3	75.6	72.4	61.4

Bibtex

@misc{guo2024instruction,
      title={Instruction Fusion: Advancing Prompt Evolution through Hybridization}, 
      author={Weidong Guo and Jiuding Yang and Kaitong Yang and Xiangyang Li and Zhuwei Rao and Yu Xu and Di Niu},
      year={2024},
      eprint={2312.15692},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}