README.md

metadata

license: bigcode-openrail-m
datasets:
  - bigcode/guanaco-commits
metrics:
  - code_eval
library_name: peft
tags:
  - code

Astraios: Parameter-Efficient Instruction Tuning Code Large Language Models

Table of Contents

1. Model Summary
2. Use
3. Training
4. Citation

Model Summary

Astraios-3B-FFT is an instruction tuned model with 15.5B parameters created by finetuning StarCoderBase on CommitPackFT & OASST as described in the Astraios paper.

• Repository: bigcode-project/astraios
• Paper: Astraios: Parameter-Efficient Instruction Tuning Code Large Language Models
• Languages: 80+ Programming languages
• ✨Astraios:

  Data	CommitPackFT+OASST	Filtered version of CommitPack and OASST for high-quality commit messages that resemble instructions
  Model	Astraios-1B	Collection of StarCoderBase-1B models instruction tuned on CommitPackFT + OASST with different tuning methods
  	Astraios-3B	Collection of StarCoderBase-3B (3B parameters) models instruction tuned on CommitPackFT + OASST with different tuning methods
  	Astraios-7B	Collection of StarCoderBase-7B (7B parameters) models instruction tuned on CommitPackFT + OASST with different tuning methods
  	Astraios-16B	Collection of StarCoderBase-16B (16B parameters) models instruction tuned on CommitPackFT + OASST with different tuning methods
  Evaluation	BigCloneBench	Dataset for clone detection; We use 2,000 samples for evaluation
  	Devign	Dataset for defect detection; We use 2,000 samples for evaluation
  	HumanEvalPack	Extension of OpenAI's HumanEval to cover 3 scenarios across 6 languages
  	ReCode	Dataset for the robustness of code generation, covering 4 variants
  	Asleep At The Keyboard	Datasets for security of code generation; We use DoW for evaluation

Use

Intended use

The model follows instructions provided in the input. You should always preface your input with "Question: " and finish it with "Answer:", for example: "Question: Please write a function in Python that performs bubble sort.

Answer:"

Feel free to share your generations in the Community tab!

Generation

# pip install -q transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "bigcode/astraios-3b-fft"
model = AutoModelForCausalLM.from_pretrained(checkpoint)
device = "cuda" # for GPU usage or "cpu" for CPU usage

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)

inputs = tokenizer.encode("Question: Please write a function in Python that performs bubble sort.

Answer:", return_tensors="pt").to(device)
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

Training

Model

• Architecture: GPT-2 model with multi-query attention and Fill-in-the-Middle objective
• Steps: 250k pretraining & 200 instruction tuning
• Precision: fp32

Hardware

• Pretraining:
  • GPUs: 512 Tesla A100
  • Training time: 24 days
• Instruction tuning:
  • GPUs: 8 Tesla A100

Software

• Orchestration: Megatron-LM/Transformers
• Neural networks: PyTorch

Citation