We introduce CodeCapybara - A Code specialized Instruction-following Large Language Model. This repo also attempts to evaluate and reproduce performance results of existing LLMs for code, such as Llama, Alpaca and CodeAlpaca for code generation benchmarks (HumanEval and MBPP).
- First attempt to reproduce of LLaMA results on widely recognized Code Generation benchmarks
- CodeCapybara is fine-tuned from Llama 7B. Larger models will be available soon. You can find our checkpoints at this.
- We use our own dataset in larger scale and more diverse to fine-tune Llama under an instruction-tuning style.
- Improved evaluation results on HumanEval in comparison to LLaMA, Alpaca and CodeAlpaca.
- Full transparency with open source availability: all scripts and models are accessible to the community. We encourage you to contribute to CodeCapybara and help advance the field of code generation.
- CodeCapybara: Open Source LLaMA Model that Follow Instruction-Tuning for Code Generation.
We follow several recent techniques of instruction tuning to collect data and train an instruction-following model with ability to generate executable code from human language description.
We can divide our process for training CodeyCapybara into two stages:
- Data Collection: We collect data generated through OpenAI
gpt-3.5-turbo
as well as code generation supervised dataset. - Instruction Tuning: We fine-tune our model from MetaAI's LLaMA checkpoint with parameter-efficient fine-tuning methods.
In this stage, we follow previous works to collect instruction data. To ensure the quality of the code data used in the fine-tuning stage, we make some modifications from data Self-Instruct data generation procedure.
To ensure the code quality for later use as targets in the fine-tuning step, we leverage an unsupervised dataset that only contains code snippets crawled from open-sources. We then design a prompt to ask gpt-3.5-turbo
to generate a corresponding instruction for each code snippet. In other words, to obtain a pair (instruction-output), we ask gpt-3.5-turbo
to generate the instruction given the output as human written code snippet.
Our unsupervised dataset contains code functions that covers a wide range of programming problem in 10 programming languages, i.e Python, Javascript, Java, Golang, Ruby, Rust, PHP, C, C++, C#
We obtain our dataset through gpt-3.5-turbo
OpenAI API. Each instruction-output pair is generated through 2 rounds of API calling.
-
In 1st round, we include a code function (i.e output) in the prompt, and ask
gpt-3.5-turbo
to generate a corresponding instruction. -
In 2nd round, since the code function does not guarantee an executable program, we include both 1st round generated instruction and code function to a new prompt and ask the model to generate an executable program with libraries imported and dependencies implementation along with the given code function.
-
Our prompt template can be found here.
-
Our script for 2 rounds of data generation can be found here.
For the second source of data, our intention is to follow Self-Instruct paper to completely generate various code problems in the format of (Instruction-Input-Output) data from a seed dataset.
We reuse the generated instruction data from Code Alpaca to reduce API calling cost since what they did is similar to our purpose.
We also leverage the supervised code generation dataset. There are various code generation dataset with high quality and quantity, such as APPS (5,000 problems in train split), MBPP (500 problems in train split).
In this version, we select DeepMind's Code Contests dataset, which contains competitive programming problems with detailed description and test cases. The train split we employ to fine-tune our model contains 13,328 problems which results in 51,766 instruction-output pairs.
We tried 2 approaches to fine-tune LLaMA-7B checkpoint on the collected data, including:
- Full-parameter Fine-tuning
- Parameter-efficient Fine-tuning with HuggingFace's PEFT
Please refer to Checkpoint Release section for accessing to our checkpoints.
We evaluate our models as well as reproduce other models' results on 2 benchmarks, HumanEval and MBPP. All numbers are reported in zero-shot settings.
Model | Base checkpoint | pass@1 | pass@10 | pass@100 |
---|---|---|---|---|
LLaMA | decapoda-research/llama-7b-hf | 10.70 | 13.29 | 13.41 |
LLaMA | huggyllama/llama-7b | 9.7 | 12.66 | 12.80 |
Alpaca-LoRA | decapoda-research/llama-7b-hf | 8.00 | 10.00 | 10.37 |
CodeCapybara-LoRA | decapoda-research/llama-7b-hf | 9.61 | 11.62 | 12.02 |
CodeCapybara | huggyllama/llama-7b | 11.10 | 13.33 | 13.41 |
We release our data as well as other data sources used for training our models
- Our Instruction Only Generation data
- Code Apaca data
- Deepmind's CodeContests hosted on HuggingFace
We release our checkpoints hosted on HuggingFace
- CodeCapybara - Full-parameter Fine-tuning
- CodeCapypara-LoRA - Parameter-efficient Fine-tuning
conda create -n codecapybara -y
conda activate codecapybara
conda install pip -y
pip install -r requirements.txt
Let's define a function to convert instruction
and input
into a single prompt as input to our model.generate
def generate_prompt(instruction, input=None):
# Templates used by Stanford Alpaca: https://github.com/tatsu-lab/stanford_alpaca
if input is not None:
prompt = f"Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:"
else:
prompt = f"prompt_no_input": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:"
return prompt
You can choose to load full-parameter CodeCapybara
or CodeCapybara-LoRA
import sys
import torch
from transformers import LlamaTokenizer, LlamaForCausalLM
tokenizer = LlamaTokenizer.from_pretrained("Fsoft-AIC/CodeCapybara")
model = LlamaForCausalLM.from_pretrain("Fsoft-AIC/CodeCapybara",
load_in_8bit=True,
dtype=torch.float16,
device_map="auto")
model.config.pad_token_id = tokenizer.pad_token_id = 0
model.config.bos_token_id = 1
model.config.eos_token_id = 2
model.eval()
if torch.__version__ >= "2" and sys.platform != "win32":
model = torch.compile(model)
import sys
import torch
from transformers import LlamaTokenizer, LlamaForCausalLM
from peft import PeftModel
tokenizer = LlamaTokenizer.from_pretrained("decapoda-research/llama-7b-hf")
model = LlamaForCausalLM.from_pretrained("decapoda-research/llama-7b-hf",
load_in_8bit=True,
dtype=torch.float16,
device_map="auto")
model = PeftModel.from_pretrained("Fsoft-AIC/CodeCapybara-LoRA",
load_in_8bit=True,
dtype=torch.float16,
device_map="auto")
model.config.pad_token_id = tokenizer.pad_token_id = 0
model.config.bos_token_id = 1
model.config.eos_token_id = 2
model.eval()
if torch.__version__ >= "2" and sys.platform != "win32":
model = torch.compile(model)
After loading model to your device, add the following script to generate prediction
instruction = "Write a Python program that prints the first 10 Fibonacci numbers"
prompt = generate_prompt(instruction)
input_ids = tokenizer(prompt)["input_ids"]
generation_config = GenerationConfig(temperature=0.1,
top_k=40,
top_p=0.75)
with torch.no_grad():
output_ids = model.generate(inputs,
generation_config=generation_config,
max_new_tokens=128)
output = tokenizer.decode(output_ids, skip_special_tokens=True, ignore_tokenization_space=True)
print(output)
We support 2 settings to fine-tune LLaMA models. In the first setting, we refine all the parameters using Fully Sharded Data Parallel, and for the rest, we currently utilize LoRA to adapt the models to the instruction tuning task. You can easily run such settings by the command
bash scripts/train.sh
which calls main/train.py
. We also provide some arguments to customize the training process
- --train-batch-size: batch-size of each gpu for training
- --val-batch-size: batch-size of each gpu for validating
- --num-workers: number of workers in the DataLoader
- --config-path: the path of the configuration file. We provide a template in the folder
configs
- --model-type: setting's used to fine-tune. There are 2 valid values:
fine-tunning
andlora
. - --use-wandb: 0 if you don't use wandb for logging; otherwise, wandb is used.
Moreover, you can edit the configuration file
configs/config.yml
which contains some notable fields: - checkpoint
- dir: the folder contains all the checkpoints
- old_checkpoint: the path of the old checkpoint. If it is null, the model'll train from scratch; otherwise, it continues training from this checkpoint.
- epochs: the number of epochs between 2 consecutive model saves.
- epochs: number of epochs for training
- model:
- hf_model: LLaMA model in HuggingFace format
- lora: settings for LoRA method
- optimizer: specify optimizer
- scheduler: configurate the hypermeters for a warm-up learning-rate schedule
- max-seq-length: maximum length of the instruction and the response.
To evaluate checkpoints on HumanEval or MBPP benchmark, navigate to main/
cd main/
We use nucleus sampling for sampling next-token in each prediction step to generate multiple difference code outputs for each problem. Hyperparameter configuration used for our evaluation is specified in the command below.
The first part of the below command generates multiple .jsonl
files, which will be saved into path/to/prediction/directory
by inference the model. The command follows after taking predictions as input to calculate pass@k.
# model inference
export CUDA_VISIBLE_DEVICES=0,1
N_PROCS=$(echo $CUDA_VISIBLE_DEVICES | tr "," "\n" | wc -l)
NUM_ITERATIONS=10
for _ in $(seq $NUM_ITERATIONS);
do
python -m torch.distributed.run --nprocs ${N_PROCS} generate.py \
--output_dir path/to/prediction/directory \
--dataset_name 'humaneval' \
--base_model 'Fsoft-AIC/CodeCapybara' \
--lora_weights '' \
--batch_size 1 \
--num_return_sequences 20 \
--load_8bit True \
--temperature 0.1 \
--top_p 0.75 \
--top_k 40
done
# Calculating pass@k with k=1,10,100
python eval_humaneval.py --prediction_dir path/to/prediction/directory
n = NUM_ITERATIONS * batch_size * num_return_sequences
, where n
is used to estimate pass@k
as in the Codex paper.
$${pass@k} = \underset{\text { Problems }}{\mathbb{E}}\left[1-\frac{C^{k}{n-c}}{C^{k}{n}}\right]$$
Here we choose n = 200
as employed in the paper, which results in
NUM_ITERATIONS=10
batch_size=1
num_return_sequences=20
Replacing the humaneval
by mbpp
# model inference
export CUDA_VISIBLE_DEVICES=0,1
N_PROCS=$(echo $CUDA_VISIBLE_DEVICES | tr "," "\n" | wc -l)
NUM_ITERATIONS=10
for _ in $(seq $NUM_ITERATIONS);
do
python -m torch.distributed.run --nprocs ${N_PROCS} generate.py \
--output_dir path/to/prediction/directory \
--dataset_name 'mbpp' \
--base_model 'Fsoft-AIC/CodeCapybara' \
--lora_weights '' \
--batch_size 1 \
--num_return_sequences 20 \
--load_8bit True \
--temperature 0.1 \
--top_p 0.75 \
--top_k 40
done
# Calculating pass@k with k=1,10,80,100
python eval_mbpp.py --prediction_dir path/to/prediction/directory
Since MetaAI released their official LLaMA checkpoints, there have been questions and efforts on reproducing their results on HumanEval and MBPP reported in paper. This repo wishes to reproduce LLaMA and other LLMs results on widely recognized Code Generation benchmarks.
To evaluate a HuggingFace LLaMA checkpoint on HumanEval or MBPP, please pass the values of --base_model
and --dataset_name
the corresponding model and benchmark in the evaluation script example.
You can also tweak hyperparameters i.e temperature
, top-p
, top-k
for trade-off between accuracy and diversity and in prediction. Tuning hyperparameters will lead to change in final results. Community is welcome for seeking optimal hyperparameter values.
We are in our progress of evaluating LLaMA official checkpoints without HuggingFace format checkpoint conversion.
Feel free to cite us
@misc{codecapybara,
title = {CodeCapybara: Code Instruction Tuning},
author = {},
year = {2023},
}