Code Alpaca: An Instruction-following LLaMA Model trained on code generation instructions

This is the repo for the Code Alpaca project, which aims to build and share an instruction-following LLaMA model for code generation. This repo is fully based on Stanford Alpaca ,and only changes the data used for training. Training approach is the same.

The repo contains:

• The 20K data used for fine-tuning the model
• The code for generating the data
• The code for fine-tuning the model

Demo for the model can be found https://code-alpaca-demo.vercel.app/

Overview

The Code Alpaca models are fine-tuned from a 7B and 13B LLaMA model on 20K instruction-following data generated by the techniques in the Self-Instruct [1] paper, with some modifications that we discuss in the next section. Evals are still a todo.

The model is not finetuned to be safe and harmless, so be cautious.

Current release contains the data generation procedure, dataset, and training code. Model weights aren't part of the release for now, to respect OpenAI TOS and LLaMA license.

[1]: Self-Instruct: Aligning Language Model with Self Generated Instructions. Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, Hannaneh Hajishirzi. https://arxiv.org/abs/2212.10560

Data Release

data/code_alpaca_20k.json contains 20K instruction-following data used for fine-tuning the Code Alpaca model. This JSON file is a list of dictionaries, each dictionary contains the following fields:

• instruction: str, describes the task the model should perform. Each of the 20K instructions is unique.
• input: str, optional context or input for the task. For example, when the instruction is "Amend the following SQL query to select distinct elements", the input is the SQL query. Around 40% of the examples have an input.
• output: str, the answer to the instruction as generated by text-davinci-003.

We used the following prompts for fine-tuning the model:

• for examples with a non-empty input field:

Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Input:
{input}

### Response:

• for examples with an empty input field:

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

### Instruction:
{instruction}

### Response:

During inference (eg for the web demo), we use the user instruction with an empty input field (second option).

Data Generation Process

This produced an instruction-following dataset with 20K examples obtained at a much lower cost (less than $200). Also including a smaller 2k samples dataset which was used to derisk the approach and quality of the model.

Fine-tuning

Finetuned the models using standard Hugging Face training code and deepspeed with the following hyperparameters:

Given Hugging Face hasn't officially supported the LLaMA models, we fine-tuned LLaMA with Hugging Face's transformers library by installing it from a particular fork (i.e. this PR to be merged). The hash of the specific commit we installed was 68d640f7c368bcaaaecfc678f11908ebbd3d6176.

The code runs on a 8xA100 80GB, but can also run on 8xA10040GB or 4xA100 with lower batch size and gradient accumulation steps. To get the GPUs, I suggest using Lambda Labs, best pricing for the best hardware.

To reproduce the fine-tuning runs for LLaMA, first install the requirements

pip install -r requirements.txt

Then, install the particular fork of Hugging Face's transformers library.

Below is a command that fine-tunes LLaMA-7B with our dataset on a machine with 4 A100 80G GPUs using deepspeed.

Replace <your_random_port> with a port of your own, <your_path_to_hf_converted_llama_ckpt_and_tokenizer> with the path to your converted checkpoint and tokenizer (following instructions in the PR), and <your_output_dir> with where you want to store your outputs.

torchrun --nproc_per_node=8 --master_port=<your_random_port> train.py \
    --model_name_or_path <your_path_to_hf_converted_llama_ckpt_and_tokenizer>
    --data_path ./data/code_alpaca_20k.json \
    --fp16 True \
    --output_dir <your_output_dir> \
    --num_train_epochs 3 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 8 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 500 \
    --save_total_limit 1 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --deepspeed ds_config.json
    --tf32 False

Note the given training script is meant to be simple and easy to use, and is not particularly optimized.

For convenience I have included the convert_to_hf.py to covnert llama checkpoints to huggingface compatible checkpoints. (This file is taken from the hugginface transformers repo)

Citation

Cite this repo if you want to, or don't, both are fine.

@misc{codealpaca,
  author = {Sahil Chaudhary},
  title = {Code Alpaca: An Instruction-following LLaMA model for code generation},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/sahil280114/codealpaca}},
}

Naturally, you should also cite the original LLaMA paper [1] and the Self-Instruct paper [2] and the Stanford Alpaca repo.