README.md

---
license: bigcode-openrail-m
datasets:
- bigcode/guanaco-commits
metrics:
- code_eval
library_name: peft
tags:
- code
---
# Astraios: A Recipe for Parameter-Efficient Instruction Tuning Code 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](#model-summary)
2. [Use](#use)
3. [Training](#training)
4. [Citation](#citation)

# Model Summary

> Astraios-P-Tuning 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](https://github.com/bigcode-project/astraios)
- **Paper:** [Astraios: A Recipe for Parameter Efficient Instruction Tuning Code 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
```python
# 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-ptuning"
checkpoint = "bigcode/starcoderbase"
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](https://github.com/bigcode-project/octopack#training)
- **Neural networks:** [PyTorch](https://github.com/pytorch/pytorch)

# Citation

```bibtex
```