metadata

library_name: transformers
tags:
  - code-generation
  - C-programming
  - LoRA

Model Card for Hend-10/MistralActia

This model is a fine-tuned version of the Mistral-7B model, adapted for generating C programming code. It leverages LoRA for improved performance in producing accurate and contextually relevant code snippets.

Model Details

Model Description

This model is fine-tuned using the Mistral-7B architecture with LoRA to enhance its capability in generating C programming code. The fine-tuning process involved training on a dataset containing C programming tasks to improve the model's ability to generate and assist with code-related queries.

Developed by: [Hend Amri]
Model type: Language Model (Causal LM)
Finetuned from model: Mistral-7B

Model Sources [optional]

Repository: Hugging Face Model Repository

Uses

Direct Use

This model can be used directly for generating code snippets, examples, and explanations in C programming. It is useful for developers, educators, and researchers who need assistance with C code generation.

Downstream Use [optional]

The model can be integrated into applications and tools that assist with coding tasks, such as IDE plugins, coding assistants, and educational platforms for programming.

Out-of-Scope Use

The model is not suitable for generating code in languages other than C or for tasks requiring domain-specific knowledge beyond general programming.

Bias, Risks, and Limitations

Bias

The model may exhibit biases present in the training data, such as preferences for certain coding styles or conventions. Users should review generated code for accuracy and appropriateness.

Limitations

The model may struggle with highly specialized or non-standard C programming tasks. It is also limited by the quality and diversity of the training data.

Risks

The model might generate code that is syntactically correct but functionally incorrect or insecure. Always validate and test the generated code thoroughly.

Recommendations

Users should verify and test all code generated by the model. It is recommended to use the model as an aid rather than a sole source of coding solutions.

How to Get Started with the Model

Use the following code snippet to get started with the model:

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Hend-10/MistralActia"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = "Write a C program that defines an enum with bit flags and shows how to use bitwise operators to manipulate these flags."
inputs = tokenizer(prompt, return_tensors="pt")

# Generate text
outputs = model.generate(
    **inputs,
    max_length=len(inputs['input_ids'][0]) + 200,  # Ensure the generated text doesn't exceed a certain length
    eos_token_id=tokenizer.eos_token_id  # Ensure the generation stops at the end of the sequence
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Data

The training data for this model consists of a dataset formatted as JSON Lines (JSONL). It includes diverse C programming prompts and responses that have been used to fine-tune the model for generating C code. The dataset is derived from various sources related to programming tasks and examples.

Training Procedure

Preprocessing

The data preprocessing involved:

Loading the data from an Excel file and removing duplicates.
Converting the cleaned data into JSONL format suitable for the datasets library.
Formatting each entry to fit the input-output structure required for training.

Training Hyperparameters

Training regime: fp16 mixed precision (for efficiency and faster training)
Batch Size: 2 per device
Number of Epochs: 7
Learning Rate: 2e-4
Evaluation Strategy: Evaluate every 500 steps
Save Strategy: Save every 500 steps
Logging Steps: Log every 500 steps
Gradient Accumulation Steps: 8
Optimizer: paged_adamw_32bit
Weight Decay: 0.001
Gradient Clipping Threshold: 0.3
Warmup Ratio: 0.03
LR Scheduler Type: Constant

Speeds, Sizes, Times

Training Time: Approximately 53 minutes and 1 second (3276.27 seconds) for the training process.
Checkpoint Size: 16.7 MB

Evaluation

Testing Data, Factors & Metrics

Testing Data

The model was evaluated using a held-out test set derived from the same source as the training data. This set was not seen by the model during training and is used to assess the model's performance on unseen prompts.

Factors

Evaluation factors include:

Prompt complexity
Response relevance
Code correctness

Metrics

Perplexity: Measures how well the model predicts the next token.
Accuracy: Checks if the generated code is syntactically correct and fulfills the prompt requirements.

Results

The model demonstrates strong performance in generating coherent and contextually appropriate C code snippets. However, its effectiveness can vary based on the complexity and specificity of the prompts.

Summary

The model performs well on a range of C programming tasks, though further fine-tuning and evaluation on more diverse datasets could enhance its robustness and applicability.

Model Examination [optional]

The model's interpretability is an ongoing area of research. Examination of model outputs includes analyzing the consistency and relevance of generated code snippets relative to the provided prompts.

Environmental Impact

Technical Specifications [optional]

Model Architecture and Objective

The model is based on the Mistral-7B architecture, designed for generating and completing text. It has been fine-tuned to generate C programming code, focusing on providing syntactically and semantically accurate outputs.

Compute Infrastructure

The model was trained using high-performance computing infrastructure provided by [Insert Cloud Provider or Hardware Details].

Hardware

Type: GPUs/TPUs
Model: Google Colab L4 GPU, Google Colab A100

Software

Frameworks: Transformers, PyTorch
Libraries: Datasets, Accelerate, PEFT, TRL

Citation [optional]

BibTeX:

@misc{hend2024mistralactia,
  author = {Amri Hend},
  title = {MistralActia: A Fine-Tuned Mistral Model for C Code Generation},
  year = {2024},
  url = {https://huggingface.co/Hend-10/MistralActia}
}

APA:

Hend, A. (2024). MistralActia: A Fine-Tuned Mistral Model for C Code Generation. Retrieved from https://huggingface.co/Hend-10/MistralActia

Glossary

Perplexity: A measure of how well a probabilistic model predicts a sample.
FP16 Mixed Precision: A type of precision that uses 16-bit floating point numbers to improve performance and reduce memory usage.

More Information

For more details about the model and to explore my other projects, visit my GitHub profile.

Model Card Authors

Author: Amri Hend

Model Card Contact

For further inquiries, please contact hend.amri@esprit.tn or reach out via the Hugging Face Model Hub.