astraios-3b-ia3

Astraios: Parameter-Efficient Instruction Tuning Code Large Language Models

<p align="center" width="100%"> <a ><img src="https://github.com/bigcode-project/astraios/blob/main/visuals/banner.png?raw=true" alt="Astraios" style="width: 20%; min-width: 300px; display: block; margin: auto;"></a> </p>

Table of Contents

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

Model Summary

Astraios-3B-IA3 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:

<table> <tr> <th>Data</t> <td><a href=https://huggingface.co/datasets/bigcode/guanaco-commits>CommitPackFT+OASST</a></td> <td>Filtered version of CommitPack and OASST for high-quality commit messages that resemble instructions</td> </tr> <tr> <th>Model</t> <td><a href=https://huggingface.co/collections/bigcode/astraios-1b-6576ff1b8e449026ae327c1c>Astraios-1B</a></td> <td>Collection of StarCoderBase-1B models instruction tuned on CommitPackFT + OASST with different tuning methods</td> </tr> <tr> <th></t> <td><a href=https://huggingface.co/collections/bigcode/astraios-3b-6577127317ee44ff547252d3>Astraios-3B</a></td> <td>Collection of StarCoderBase-3B (3B parameters) models instruction tuned on CommitPackFT + OASST with different tuning methods</td> </tr> <tr> <th></t> <td><a href=https://huggingface.co/collections/starpeft/starcoderbase-7b-650c1f028b45cfec8e72c265>Astraios-7B</a></td> <td>Collection of StarCoderBase-7B (7B parameters) models instruction tuned on CommitPackFT + OASST with different tuning methods</td> </tr> <tr> <th></t> <td><a href=https://huggingface.co/collections/bigcode/astraios-16b-65788b7476b6de79781054cc>Astraios-16B</a></td> <td>Collection of StarCoderBase-16B (16B parameters) models instruction tuned on CommitPackFT + OASST with different tuning methods</td> </tr> <tr> <th>Evaluation</t> <td><a href=https://huggingface.co/datasets/code_x_glue_cc_clone_detection_big_clone_bench>BigCloneBench</a></td> <td>Dataset for clone detection; We use 2,000 samples for evaluation</td> </tr> <tr> <th></t> <td><a href=https://huggingface.co/datasets/code_x_glue_cc_defect_detection>Devign</a></td> <td>Dataset for defect detection; We use 2,000 samples for evaluation</td> </tr> <tr> <th></t> <td><a href=https://huggingface.co/datasets/bigcode/humanevalpack>HumanEvalPack</a></td> <td>Extension of OpenAI's HumanEval to cover 3 scenarios across 6 languages</td> </tr> <tr> <th></t> <td><a href=https://huggingface.co/datasets/RaymondLi/perturbed_humaneval>ReCode</a></td> <td>Dataset for the robustness of code generation, covering 4 variants</td> </tr> <tr> <th></t> <td><a href=https://huggingface.co/datasets/moyix/asleep_keyboard>Asleep At The Keyboard</a></td> <td>Datasets for security of code generation; We use DoW for evaluation</td> </tr> </table>

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
# pip install -e git+https://github.com/bigcode-project/astraios#subdirectory=peft
from peft import PeftModel 
from transformers import AutoModelForCausalLM, AutoTokenizer

peft_checkpoint = "bigcode/astraios-3b-ia3"
checkpoint = "bigcode/starcoderbase-3b"
model = AutoModelForCausalLM.from_pretrained(checkpoint)
model = PeftModel.from_pretrained(model, peft_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