Fsoft-AIC/Codebert-docstring-inconsistency

Table of Contents

• Model Description
• Model Details
• Usage
• Limitations
• Additional Information
  • Licensing Information
  • Citation Information

Model Description

This model is developed based on Codebert and a 5M subset of The Vault to detect the inconsistency between docstring/comment and function. It is used to remove noisy examples in The Vault dataset.

More information:

• Repository: FSoft-AI4Code/TheVault
• Paper: The Vault: A Comprehensive Multilingual Dataset for Advancing Code Understanding and Generation
• Contact: [email protected]

Model Details

• Developed by: Fsoft AI Center
• License: MIT
• Model type: Transformer-Encoder based Language Model
• Architecture: BERT-base
• Data set: The Vault
• Tokenizer: Byte Pair Encoding
• Vocabulary Size: 50265
• Sequence Length: 512
• Language: English and 10 Programming languages (Python, Java, JavaScript, PHP, C#, C, C++, Go, Rust, Ruby)
• Training details:
  • Self-supervised learning, binary classification
  • Positive class: Original code-docstring pair
  • Negative class: Random pairing code and docstring

Usage

The input to the model follows the below template:

"""
Template:
<s>{docstring}</s></s>{code}</s>

Example:
from transformers import AutoTokenizer

#Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("Fsoft-AIC/Codebert-docstring-inconsistency")

input = "<s>Sum two integers</s></s>def sum(a, b):\n    return a + b</s>"
tokenized_input = tokenizer(input, add_special_tokens= False)
"""

Using model with Jax and Pytorch

from transformers import AutoTokenizer, AutoModelForSequenceClassification, FlaxAutoModelForSequenceClassification

#Load model with jax
model = FlaxAutoModelForSequenceClassification.from_pretrained("Fsoft-AIC/Codebert-docstring-inconsistency")

#Load model with torch
model = AutoModelForSequenceClassification.from_pretrained("Fsoft-AIC/Codebert-docstring-inconsistency")

Limitations

This model is trained on 5M subset of The Vault in a self-supervised manner. Since the negative samples are generated artificially, the model's ability to identify instances that require a strong semantic understanding between the code and the docstring might be restricted.

It is hard to evaluate the model due to the unavailable labeled datasets. GPT-3.5-turbo is adopted as a reference to measure the correlation between the model and GPT-3.5-turbo's scores. However, the result could be influenced by GPT-3.5-turbo's potential biases and ambiguous conditions. Therefore, we recommend having human labeling dataset and fine-tune this model to achieve the best result.

Additional information

Licensing Information

MIT License

Citation Information

@article{manh2023vault,
  title={The Vault: A Comprehensive Multilingual Dataset for Advancing Code Understanding and Generation},
  author={Manh, Dung Nguyen and Hai, Nam Le and Dau, Anh TV and Nguyen, Anh Minh and Nghiem, Khanh and Guo, Jin and Bui, Nghi DQ},
  journal={arXiv preprint arXiv:2305.06156},
  year={2023}
}