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.
- 🔥🔥🔥[2024/01/04] We released WizardCoder-33B-V1.1 trained from deepseek-coder-33b-base, the SOTA OSS Code LLM on EvalPlus Leaderboard, achieves 79.9 pass@1 on HumanEval, 73.2 pass@1 on HumanEval-Plus, 78.9 pass@1 on MBPP, and 66.9 pass@1 on MBPP-Plus. WizardCoder-33B-V1.1 outperforms ChatGPT 3.5, Gemini Pro, and DeepSeek-Coder-33B-instruct on HumanEval and HumanEval-Plus pass@1. WizardCoder-33B-V1.1 is comparable with ChatGPT 3.5, and surpasses Gemini Pro on MBPP and MBPP-Plus pass@1.
- [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.
❗❗❗This performance is 100% reproducible!
❗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 | HumanEval+ | MBPP | MBPP+ |
|---|---|---|---|---|---|---|
| GPT-4-Turbo (Nov 2023) | - | - | 85.4 | 81.7 | 83.0 | 70.7 |
| GPT-4 (May 2023) | - | - | 88.4 | 76.8 | - | - |
| GPT-3.5-Turbo (Nov 2023) | - | - | 72.6 | 65.9 | 81.7 | 69.4 |
| Gemini Pro | - | - | 63.4 | 55.5 | 72.9 | 57.9 |
| DeepSeek-Coder-33B-instruct | - | - | 78.7 | 72.6 | 78.7 | 66.7 |
| WizardCoder-33B-V1.1 | 🤗 HF Link | 📃 [WizardCoder] | 79.9 | 73.2 | 78.9 | 66.9 |
| WizardCoder-Python-34B-V1.0 | 🤗 HF Link | 📃 [WizardCoder] | 73.2 | 64.6 | 73.2 | 59.9 |
| WizardCoder-15B-V1.0 | 🤗 HF Link | 📃 [WizardCoder] | 59.8 | 52.4 | -- | -- |
| WizardCoder-Python-13B-V1.0 | 🤗 HF Link | 📃 [WizardCoder] | 64.0 | -- | -- | -- |
| WizardCoder-Python-7B-V1.0 | 🤗 HF Link | 📃 [WizardCoder] | 55.5 | -- | -- | -- |
| WizardCoder-3B-V1.0 | 🤗 HF Link | 📃 [WizardCoder] | 34.8 | -- | -- | -- |
| WizardCoder-1B-V1.0 | 🤗 HF Link | 📃 [WizardCoder] | 23.8 | -- | -- | -- |
- 📣 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.
🔥 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).
🔥 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.
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).
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.
Thanks to the enthusiastic friends, their video introductions are more lively and interesting.
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.)
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:
- According to the instructions of Llama-X, install the environment, download the training code, and deploy. (Note:
deepspeed==0.9.2andtransformers==4.29.2) - Replace the
train.pywith thetrain_wizardcoder.pyin our repo (src/train_wizardcoder.py) - Login Huggingface:
huggingface-cli login- 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 TrueWe 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 jsonlinesThe 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:
- According to the instructions of HumanEval, install the environment.
- 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 ${output_path} --num_gpus 4- Run the post processing code
src/process_humaneval.pyto collect the code completions from all answer files.
output_path=preds/T${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 ${output_path}.jsonl❗❗❗This performance is 100% reproducible!
We also provide all generated results in WizardLM/WizardCoder/data/humaneval_mbpp_wizardcoder33b_v1.1_results.zip
transformers==4.36.2
vllm==0.2.5
(1) HumanEval and HumanEval-Plus
- Step 1
Code Generation (w/o accelerate)
model="WizardLM/WizardCoder-33B-V1.1"
temp=0.0
max_len=2048
pred_num=1
num_seqs_per_iter=1
output_path=preds/T${temp}_N${pred_num}_WizardCoder-33B-V1.1_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 % $gpu_num == 0)); then wait; fi
doneCode Generation (w/ vllm accelerate)
model="WizardLM/WizardCoder-33B-V1.1"
temp=0.0
max_len=2048
pred_num=1
num_seqs_per_iter=1
output_path=preds/T${temp}_N${pred_num}_WizardCoder-33B-V1.1_Greedy_Decode_vllm
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 ${output_path} --num_gpus 4 --overwrite- Step 2: Get the score
Install Eval-Plus benchmark.
git clone https://github.com/evalplus/evalplus.git
cd evalplus
export PYTHONPATH=$PYTHONPATH:$(pwd)
pip install -r requirements.txtGet HumanEval and HumanEval-Plus scores.
output_path=preds/T0.0_N1_WizardCoder-33B-V1.1_Greedy_Decode
echo 'Output path: '$output_path
python process_humaneval.py --path ${output_path} --out_path ${output_path}.jsonl --add_prompt
evalplus.evaluate --dataset humaneval --samples ${output_path}.jsonl(2) MBPP and MBPP-Plus
The preprocessed questions are provided in WizardLM/WizardCoder/data/mbppplus.json
- Step 1
Code Generation (w/o accelerate)
model="WizardLM/WizardCoder-33B-V1.1"
temp=0.0
max_len=2048
pred_num=1
num_seqs_per_iter=1
output_path=preds/MBPP_T${temp}_N${pred_num}_WizardCoder-33B-V1.1_Greedy_Decode
mkdir -p ${output_path}
echo 'Output path: '$output_path
echo 'Model to eval: '$model
# 399 problems, 50 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 mbppplus_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 ${output_path} --mbpp_path "mbppplus.json" --greedy_decode
) &
if (($index % $gpu_num == 0)); then wait; fi
doneCode Generation (w/ vllm accelerate)
model="WizardLM/WizardCoder-33B-V1.1"
temp=0.0
max_len=2048
pred_num=1
num_seqs_per_iter=1
output_path=preds/MBPP_T${temp}_N${pred_num}_WizardCoder-33B-V1.1_Greedy_Decode_vllm
mkdir -p ${output_path}
echo 'Output path: '$output_path
echo 'Model to eval: '$model
CUDA_VISIBLE_DEVICES=0,1,2,3 python mbppplus_gen_vllm.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 ${output_path} --mbpp_path "mbppplus.json" --num_gpus 4- Step 2: Get the score
Install Eval-Plus benchmark.
git clone https://github.com/evalplus/evalplus.git
cd evalplus
export PYTHONPATH=$PYTHONPATH:$(pwd)
pip install -r requirements.txtGet HumanEval and HumanEval-Plus scores.
output_path=preds/MBPP_T0.0_N1_WizardCoder-33B-V1.1_Greedy_Decode
echo 'Output path: '$output_path
python mbppplus_process_preds.py --path ${output_path} --out_path ${output_path}.jsonl --add_prompt
evalplus.evaluate --dataset mbpp --samples ${output_path}.jsonl❗❗❗This performance is 100% reproducible!
- Step 1: Setup the environment
conda create -n eval python=3.10
conda activate eval
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.3 -c pytorch- Step 2: Install the packages
transformers==4.31.0
numpy
fire
sentencepiece
deepspeed==0.10.0
accelerate
vllm==0.1.4
pandas
ray
pyarrow
- Step 3: Install Human-Eval from OpenAI
git clone https://github.com/openai/human-eval.git
pip install -e human-evaluncomment the execution call in human-eval/human_eval/execution.py
- Step 4: Generate Answer
use the code WizardLM/blob/main/WizardCoder/src/humaneval_gen_vllm.py to generate answer.
model="WizardLM/WizardCoder-Python-34B-V1.0"
temp=0.0
max_len=2048
pred_num=1
num_seqs_per_iter=1
output_path=preds/T${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 ${output_path} --num_gpus 4- Step 5: Get the score
use the code WizardLM/blob/main/WizardCoder/src/process_humaneval.py to get the score.
output_path=preds/T0.0_N1
echo 'Output path: '$output_path
python process_humaneval.py --path ${output_path} --out_path ${output_path}.jsonl --add_prompt
evaluate_functional_correctness ${output_path}.jsonl❗❗❗This performance is 100% reproducible!
Run the following script to generate the answer with greedy decoding. Then follow the above steps 2 and 3 to get the evaluation result.
❗We also provide the generated codes in data/humaneval.59.8.gen.zip
model="WizardLM/WizardCoder-15B-V1.0"
temp=0.0
max_len=2048
pred_num=1
num_seqs_per_iter=1
output_path=preds/T${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 % $gpu_num == 0)); then wait; fi
done- Run the following script to generate the answer.
model="/path/to/your/model"
temp=0.2
max_len=2048
pred_num=200
num_seqs_per_iter=2
output_path=preds/MBPP_T${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 % $gpu_num == 0)); then wait; fi
done- Run the post processing code
src/process_mbpp.pyto collect the code completions from all answer files.
output_path=preds/MBPP_T${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 ${mbpp_path} --add_prompt- Evaluate the
MBPP_T${temp}_N${pred_num}.jsonlwith bigcode-evaluation-harness.
Acknowledgement: The evaluation code humaneval_gen.py, mbpp_gen.py and bash scripts are modified from the great works of CodeT5.
Please cite the repo if you use the data or code in this repo.
@article{luo2023wizardcoder,
title={WizardCoder: Empowering Code Large Language Models with Evol-Instruct},
author={Luo, Ziyang and Xu, Can and Zhao, Pu and Sun, Qingfeng and Geng, Xiubo and Hu, Wenxiang and Tao, Chongyang and Ma, Jing and Lin, Qingwei and Jiang, Daxin},
journal={arXiv preprint arXiv:2306.08568},
year={2023}
}
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.