BSCode-Optimizer is an advanced AI-driven tool designed for analyzing, optimizing, and fixing code inefficiencies and bugs. It leverages state-of-the-art generative models such as BSJCode-1-Stable, Code Llama, and Terjman-Ultra to provide robust support for code refinement, error detection, and multilingual explanations. This tool is ideal for developers, educators, and researchers looking to streamline their workflows and enhance code quality.
- Code Analysis: Identifies inefficiencies, bugs, and potential optimizations in code.
- Code Optimization: Rewrites and optimizes code for better performance and readability.
- Multilingual Support: Explains and translates code-related instructions in Moroccan Darija, English, and other languages.
- Customizable Pipelines: Supports BSJCode-1-Stable for structured code optimization, Code Llama for code understanding, and Terjman-Ultra for multilingual translations.
- Extensibility: Easily integrate into existing workflows via APIs or Hugging Face’s Transformers library.
The BSCode-Optimizer pipeline is built with multiple components to handle diverse requirements:
This image illustrates the high-level workflow of the BSCode-Optimizer, showing the interaction between different components and data flows.
This diagram provides a detailed view of the architecture, including how models like BSJCode-1-Stable, Code Llama, and Terjman-Ultra interact within the pipeline.
The front-end of BSCode-Optimizer is designed to provide an intuitive user experience, enabling seamless interaction with the API and models.
This diagram provides a technical overview of how different modules of BSCode-Optimizer are connected and function together.
- Python 3.8+
- CUDA-enabled GPU (optional for accelerated inference)
- Hugging Face Transformers
pip install transformers torch accelerate flask pyngrokgit clone https://github.com/your-repo/BSCode-Optimizer.git
cd BSCode-OptimizerThe BSCode-Optimizer includes a Flask API for seamless integration into workflows.
python app.pycurl -X POST http://localhost:5000/predict \
-H "Content-Type: application/json" \
-d '{
"input": "public int findMin(List<Integer> nums) { ... }",
"service": "Pro"
}'{
"output": "Optimized and fixed code output here."
}You can also interact with the models directly using the following example:
from transformers import AutoModelForCausalLM, AutoTokenizer
# Load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("BSAtlas/BSJCode-1-Stable")
model = AutoModelForCausalLM.from_pretrained("BSAtlas/BSJCode-1-Stable")
# Input code snippet
code_snippet = "public int findMin(List<Integer> nums) { ... }"
# Tokenize input
tokens = tokenizer(code_snippet, return_tensors="pt")
# Generate optimized code
output = model.generate(**tokens)
optimized_code = tokenizer.decode(output[0], skip_special_tokens=True)
print(optimized_code)from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("atlasia/Terjman-Ultra")
model = AutoModelForSeq2SeqLM.from_pretrained("atlasia/Terjman-Ultra")
# Translate code explanation to Darija
input_text = "Explain this code snippet in Moroccan Darija."
tokens = tokenizer(input_text, return_tensors="pt")
translation = model.generate(**tokens)
output_text = tokenizer.decode(translation[0], skip_special_tokens=True)
print(output_text)The development of BSCode-Optimizer was a collaborative effort led by an expert team of AI researchers, software engineers, and linguists. Key contributors include:
- Marouane Asmih
- Ahmed Elmessaoudi
- Habbach Oualid
- LamrabetChaymae
- Programming Languages: Python, Java
- Frameworks: Flask, Hugging Face Transformers
- Libraries: PyTorch, BitsAndBytes
- Hardware: NVIDIA A100 GPUs for training and inference
- Datasets: Custom-curated datasets for code and multilingual tasks
- Purpose: Java code optimization and error detection
- Key Features:
- Identifies bugs
- Provides optimized solutions with O-complexity analysis
- Training Dataset: Java-focused dataset with curated examples
- Inference Time: <200ms per code snippet on A100 GPU
- Purpose: General-purpose code synthesis and understanding
- Capabilities:
- Code completion
- Infilling
- Instruction following
- Training Period: January 2023 – July 2023
- Purpose: Multilingual code explanations and translations
- Capabilities:
- Transliteration from Moroccan Darija (Arabizi)
- Accurate Arabic-English translations
- Performance:
- Loss: 2.7070
- BLEU: 4.6998
- Gen Len: 35.6088 tokens
- Dataset: Java-specific curated dataset
- GPU: NVIDIA A100-40GB
- Training Time: 32 hours
- Base Model:
facebook/nllb-200-1.3B - Dataset: darija_english dataset with curated corpora
- Hyperparameters:
- Learning Rate:
3e-05 - Epochs:
25 - Batch Size:
16
- Learning Rate:
- BSJCode-1-Stable:
- May not handle non-Java languages effectively.
- Code Llama:
- Accuracy can vary for large, complex codebases.
- Terjman-Ultra:
- Limited by the quality of tokenizer for transliterations.
- Unpredictable Outputs: As with all LLMs, outputs may contain inaccuracies or undesirable content.
- Responsible Use: Follow Meta’s Responsible Use Guide.
- Privacy: Avoid using sensitive or proprietary code as input.
- BSJCode-1-Stable: Custom commercial license
- Code Llama: Meta’s custom commercial license
- Terjman-Ultra: Available under Hugging Face model card terms
We welcome contributions, feedback, and suggestions to improve BSCode-Optimizer. Please open an issue or submit a pull request on the GitHub repository.
This project utilizes advanced language models and training techniques from the following:
- Meta’s Code Llama
- BSJCode-1-Stable by BSAtlas
- Terjman-Ultra for multilingual capabilities
Special thanks to the global developer community and contributors who provided invaluable feedback and support during development.