WizardCoder: Empowering Code Large Language Models with Evol-Instruct

To develop our WizardCoder model, we begin by adapting the Evol-Instruct method specifically for coding tasks. This involves tailoring the prompt to the domain of code-related instructions. Subsequently, we fine-tune the Code LLMs, StarCoder or Code LLama, utilizing the newly created instruction-following training set.

News

• 🔥🔥🔥[2023/08/26] We released WizardCoder-Python-34B-V1.0 , which achieves the 73.2 pass@1 and surpasses GPT4 (2023/03/15), ChatGPT-3.5, and Claude2 on the HumanEval Benchmarks.
• [2023/06/16] We released WizardCoder-15B-V1.0 , which achieves the 57.3 pass@1 and surpasses Claude-Plus (+6.8), Bard (+15.3) and InstructCodeT5+ (+22.3) on the HumanEval Benchmarks.

❗Note: There are two HumanEval results of GPT4 and ChatGPT-3.5. The 67.0 and 48.1 are reported by the official GPT4 Report (2023/03/15) of OpenAI. The 82.0 and 72.5 are tested by ourselves with the latest API (2023/08/26).

Model	Checkpoint	Paper	HumanEval	MBPP	Demo	License
WizardCoder-Python-34B-V1.0	🤗 HF Link	📃 [WizardCoder]	73.2	61.2	Demo	Llama2
WizardCoder-15B-V1.0	🤗 HF Link	📃 [WizardCoder]	59.8	50.6	--	OpenRAIL-M
WizardCoder-Python-13B-V1.0	🤗 HF Link	📃 [WizardCoder]	64.0	55.6	--	Llama2
WizardCoder-3B-V1.0	🤗 HF Link	📃 [WizardCoder]	34.8	37.4	--	OpenRAIL-M
WizardCoder-1B-V1.0	🤗 HF Link	📃 [WizardCoder]	23.8	28.6	--	OpenRAIL-M

• 📣 Please refer to our Twitter account https://twitter.com/WizardLM_AI and HuggingFace Repo https://huggingface.co/WizardLM . We will use them to announce any new release at the 1st time.

Comparing WizardCoder-Python-34B-V1.0 with Other LLMs.

🔥 The following figure shows that our WizardCoder-Python-34B-V1.0 attains the second position in this benchmark, surpassing GPT4 (2023/03/15, 73.2 vs. 67.0), ChatGPT-3.5 (73.2 vs. 72.5) and Claude2 (73.2 vs. 71.2).

❗❗❗Note: This performance is 100% reproducible! If you cannot reproduce it, please follow the steps in Evaluation.

❗Note: There are two HumanEval results of GPT4 and ChatGPT-3.5. The 67.0 and 48.1 are reported by the official GPT4 Report (2023/03/15) of OpenAI. The 82.0 and 72.5 are tested by ourselves with the latest API (2023/08/26).

Comparing WizardCoder-15B-V1.0 with the Closed-Source Models.

🔥 The following figure shows that our WizardCoder attains the third position in this benchmark, surpassing Claude-Plus (59.8 vs. 53.0) and Bard (59.8 vs. 44.5). Notably, our model exhibits a substantially smaller size compared to these models.

❗❗❗Note: This performance is 100% reproducible! If you cannot reproduce it, please follow the steps in Evaluation.

❗Note: In this study, we copy the scores for HumanEval and HumanEval+ from the LLM-Humaneval-Benchmarks. Notably, all the mentioned models generate code solutions for each problem utilizing a single attempt, and the resulting pass rate percentage is reported. Our WizardCoder generates answers using greedy decoding and tests with the same code.

Comparing WizardCoder-15B-V1.0 with the Open-Source Models.

The following table clearly demonstrates that our WizardCoder exhibits a substantial performance advantage over all the open-source models. ❗If you are confused with the different scores of our model (57.3 and 59.8), please check the Notes.

Model	HumanEval Pass@1	MBPP Pass@1
CodeGen-16B-Multi	18.3	20.9
CodeGeeX	22.9	24.4
LLaMA-33B	21.7	30.2
LLaMA-65B	23.7	37.7
PaLM-540B	26.2	36.8
PaLM-Coder-540B	36.0	47.0
PaLM 2-S	37.6	50.0
CodeGen-16B-Mono	29.3	35.3
Code-Cushman-001	33.5	45.9
StarCoder-15B	33.6	43.6*
InstructCodeT5+	35.0	--
WizardLM-30B 1.0	37.8	--
WizardCoder-15B 1.0	57.3	51.8

❗Note: The reproduced result of StarCoder on MBPP.

❗Note: The above table conducts a comprehensive comparison of our WizardCoder with other models on the HumanEval and MBPP benchmarks. We adhere to the approach outlined in previous studies by generating 20 samples for each problem to estimate the pass@1 score and evaluate with the same code. The scores of GPT4 and GPT3.5 reported by OpenAI are 67.0 and 48.1 (maybe these are the early version GPT4&3.5).

Call for Feedbacks

We welcome everyone to use your professional and difficult instructions to evaluate WizardCoder, and show us examples of poor performance and your suggestions in the issue discussion area. We are focusing on improving the Evol-Instruct now and hope to relieve existing weaknesses and issues in the the next version of WizardCoder. After that, we will open the code and pipeline of up-to-date Evol-Instruct algorithm and work with you together to improve it.

Unofficial Video Introductions

Thanks to the enthusiastic friends, their video introductions are more lively and interesting.

1. WizardCoder AI Is The NEW ChatGPT's Coding TWIN!

Contents

1. Online Demo

2. Fine-tuning

3. Inference

4. Evaluation

5. Citation

6. Disclaimer

Online Demo

We will provide our latest models for you to try for as long as possible. If you find a link is not working, please try another one. At the same time, please try as many real-world and challenging code-related problems that you encounter in your work and life as possible. We will continue to evolve our models with your feedbacks.

Demo Link (We adopt the greedy decoding now.)

Fine-tuning

We fine-tune WizardCoder using the modified code train.py from Llama-X. We fine-tune StarCoder-15B with the following hyperparameters:

Hyperparameter	StarCoder-15B
Batch size	512
Learning rate	2e-5
Epochs	3
Max length	2048
Warmup step	30
LR scheduler	cosine

To reproduce our fine-tuning of WizardCoder, please follow the following steps:

1. According to the instructions of Llama-X, install the environment, download the training code, and deploy. (Note: deepspeed==0.9.2 and transformers==4.29.2)
2. Replace the train.py with the train_wizardcoder.py in our repo (src/train_wizardcoder.py)
3. Login Huggingface:

huggingface-cli login

4. Execute the following training command:

deepspeed train_wizardcoder.py \
    --model_name_or_path "bigcode/starcoder" \
    --data_path "/your/path/to/code_instruction_data.json" \
    --output_dir "/your/path/to/ckpt" \
    --num_train_epochs 3 \
    --model_max_length 2048 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 50 \
    --save_total_limit 2 \
    --learning_rate 2e-5 \
    --warmup_steps 30 \
    --logging_steps 2 \
    --lr_scheduler_type "cosine" \
    --report_to "tensorboard" \
    --gradient_checkpointing True \
    --deepspeed configs/deepspeed_config.json \
    --fp16 True

Inference

We provide the decoding script for WizardCoder, which reads a input file and generates corresponding responses for each sample, and finally consolidates them into an output file.

You can specify base_model, input_data_path and output_data_path in src\inference_wizardcoder.py to set the decoding model, path of input file and path of output file.

pip install jsonlines

The decoding command is:

python src\inference_wizardcoder.py \
    --base_model "/your/path/to/ckpt" \
    --input_data_path "/your/path/to/input/data.jsonl" \
    --output_data_path "/your/path/to/output/result.jsonl"

The format of data.jsonl should be:

{"idx": 11, "Instruction": "Write a Python code to count 1 to 10."}
{"idx": 12, "Instruction": "Write a Java code to sum 1 to 10."}

The prompt for our WizardCoder in src\inference_wizardcoder.py is:

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Response:

Evaluation

HumanEval

1. According to the instructions of HumanEval, install the environment.
2. Run the following scripts to generate the answer.

• (1) For WizardCoder-15B-V1.0 (base on StarCoder)

model="/path/to/your/model"
temp=0.2
max_len=2048
pred_num=200
num_seqs_per_iter=2

output_path=preds/T${temp}_N${pred_num}

mkdir -p ${output_path}
echo 'Output path: '$output_path
echo 'Model to eval: '$model

# 164 problems, 21 per GPU if GPU=8
index=0
gpu_num=8
for ((i = 0; i < $gpu_num; i++)); do
  start_index=$((i * 21))
  end_index=$(((i + 1) * 21))

  gpu=$((i))
  echo 'Running process #' ${i} 'from' $start_index 'to' $end_index 'on GPU' ${gpu}
  ((index++))
  (
    CUDA_VISIBLE_DEVICES=$gpu python humaneval_gen.py --model ${model} \
      --start_index ${start_index} --end_index ${end_index} --temperature ${temp} \
      --num_seqs_per_iter ${num_seqs_per_iter} --N ${pred_num} --max_len ${max_len} --output_path ${output_path}
  ) &
  if (($index % $gpu_num == 0)); then wait; fi
done

• (2) For WizardCoder-Python-34B-V1.0 (base on CodeLLama)

pip install vllm # This can acclerate the inference process a lot.
pip install transformers==4.31.0

model="/path/to/your/model"
temp=0.2
max_len=2048
pred_num=200
num_seqs_per_iter=2

output_path=preds/T${temp}_N${pred_num}

mkdir -p ${output_path}
echo 'Output path: '$output_path
echo 'Model to eval: '$model

CUDA_VISIBLE_DEVICES=0,1,2,3 python humaneval_gen_vllm.py --model ${model} \
  --start_index 0 --end_index 164 --temperature ${temp} \
  --num_seqs_per_iter ${num_seqs_per_iter} --N ${pred_num} --max_len ${max_len} --output_path ${output_path} --num_gpus 4

3. Run the post processing code src/process_humaneval.py to collect the code completions from all answer files.

output_path=preds/T${temp}_N${pred_num}

echo 'Output path: '$output_path
python process_humaneval.py --path ${output_path} --out_path ${output_path}.jsonl --add_prompt

evaluate_functional_correctness ${output_path}.jsonl

How to Reproduce the 59.8 Pass@1 on HumanEval with Greedy Decoding?

❗❗❗This performance is 100% reproducible!

Run the following script to generate the answer with greedy decoding. Then follow the above steps 2 and 3 to get the evaluation result.

❗We also provide the generated codes in data/humaneval.59.8.gen.zip

model="WizardLM/WizardCoder-15B-V1.0"
temp=0.0
max_len=2048
pred_num=1
num_seqs_per_iter=1

output_path=preds/T${temp}_N${pred_num}_WizardCoder_Greedy_Decode

mkdir -p ${output_path}
echo 'Output path: '$output_path
echo 'Model to eval: '$model

# 164 problems, 21 per GPU if GPU=8
index=0
gpu_num=8
for ((i = 0; i < $gpu_num; i++)); do
  start_index=$((i * 21))
  end_index=$(((i + 1) * 21))

  gpu=$((i))
  echo 'Running process #' ${i} 'from' $start_index 'to' $end_index 'on GPU' ${gpu}
  ((index++))
  (
    CUDA_VISIBLE_DEVICES=$gpu python humaneval_gen.py --model ${model} \
      --start_index ${start_index} --end_index ${end_index} --temperature ${temp} \
      --num_seqs_per_iter ${num_seqs_per_iter} --N ${pred_num} --max_len ${max_len} --output_path ${output_path} --greedy_decode
  ) &
  if (($index % $gpu_num == 0)); then wait; fi
done

MBPP

1. Run the following script to generate the answer.

model="/path/to/your/model"
temp=0.2
max_len=2048
pred_num=200
num_seqs_per_iter=2

output_path=preds/MBPP_T${temp}_N${pred_num}
mbpp_path=data/mbpp.test.jsonl # we provide this file in data/mbpp.test.zip

mkdir -p ${output_path}
echo 'Output path: '$output_path
echo 'Model to eval: '$model

# 500 problems, 63 per GPU if GPU=8
index=0
gpu_num=8
for ((i = 0; i < $gpu_num; i++)); do
  start_index=$((i * 50))
  end_index=$(((i + 1) * 50))

  gpu=$((i))
  echo 'Running process #' ${i} 'from' $start_index 'to' $end_index 'on GPU' ${gpu}
  ((index++))
  (
    CUDA_VISIBLE_DEVICES=$gpu python mbpp_gen.py --model ${model} \
      --start_index ${start_index} --end_index ${end_index} --temperature ${temp} \
      --num_seqs_per_iter ${num_seqs_per_iter} --N ${pred_num} --max_len ${max_len} --output_path ${output_path} --mbpp_path ${mbpp_path}
  ) &
  if (($index % $gpu_num == 0)); then wait; fi
done

3. Run the post processing code src/process_mbpp.py to collect the code completions from all answer files.

output_path=preds/MBPP_T${temp}_N${pred_num}
mbpp_path=data/mbpp.test.jsonl # we provide this file in data/mbpp.test.zip

echo 'Output path: '$output_path
python process_mbpp.py --path ${output_path} --out_path ${output_path}.jsonl --mbpp_path ${mbpp_path} --add_prompt

4. Evaluate the MBPP_T${temp}_N${pred_num}.jsonl with bigcode-evaluation-harness.

Acknowledgement: The evaluation code humaneval_gen.py, mbpp_gen.py and bash scripts are modified from the great works of CodeT5.

Citation

Please cite the repo if you use the data or code in this repo.

@misc{luo2023wizardcoder,
      title={WizardCoder: Empowering Code Large Language Models with Evol-Instruct}, 
      author={Ziyang Luo and Can Xu and Pu Zhao and Qingfeng Sun and Xiubo Geng and Wenxiang Hu and Chongyang Tao and Jing Ma and Qingwei Lin and Daxin Jiang},
      year={2023},
      eprint={2306.08568},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Disclaimer

WizardCoder model follows the same license as StarCoder. The content produced by any version of WizardCoder is influenced by uncontrollable variables such as randomness, and therefore, the accuracy of the output cannot be guaranteed by this project. This project does not accept any legal liability for the content of the model output, nor does it assume responsibility for any losses incurred due to the use of associated resources and output results.

Star History