{}

iLLAMA: LLM for App Issue Detection and Prioritization Obtained by Fine-Tuning LLAMA 3

This repository contains a fine-tuned version of LLAMA 3 using the Unsloth framework and the vitormesaque/irisk dataset. The model is designed for detecting issues in text data.

Model Details

Developed by: Vitor Mesaque
Model type: App Issue Detection Model
Language: English
License: apache-2.0
Finetuned from model : unsloth/llama-3-8b-bnb-4bit
Datasets: vitormesaque/irisk

The vitormesaque/irisk dataset was obtained through the knowledge base of the MApp-IDEA research project.

Model Usage

How to Get Started with the Model

Use the code below to get started with the model:


# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("vitormesaque/i-llama")
model = AutoModelForCausalLM.from_pretrained("vitormesaque/i-llama")

Usage



FastLanguageModel.for_inference(model) # Enable native 2x faster inference
inputs = tokenizer(
[
    irisk_prompt.format(
        "Extract issues from the user review in JSON format. For each issue, provide label, functionality, severity (1-5), likelihood (1-5), category (Bug, User Experience, Performance, Security, Compatibility, Functionality, UI, Connectivity, Localization, Accessibility, Data Handling, Privacy, Notifications, Account Management, Payment, Content Quality, Support, Updates, Syncing, Customization), and the sentence.", # instruction
        "I used to love this app, but now it's become frustrating as hell. We can't see lyrics, we can't CHOOSE WHAT SONG WE WANT TO LISTEN TO, we can't skip a song more than a few times, there are ads after every two songs, and all in all it's a horrible overrated app. If I could give this 0 stars, I would.", # input
        "", # output - leave this blank for generation!
    )
], return_tensors = "pt").to("cuda")


from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 512)

Evaluation

The model was evaluated using a separate portion of the vitormesaque/irisk dataset.

Bias, Risks, and Limitations

While the model is effective in detecting issues, it may exhibit biases present in the training data. Users should be aware of these potential biases and consider them when interpreting results.

Recommendations

Users should conduct additional evaluations in the specific context of use to ensure reliability and fairness.

Citation

If you use this model in your research, please cite it as follows:

BibTeX:

@misc{vitormesaque2024llama3,
  author = {Vitor Mesaque Alves de Lima},
  title = {iLLAMA: LLM for App Issue Detection and Prioritization Obtained by Fine-Tuning LLAMA 3},
  year = {2024},
  url = {https://huggingface.co/vitormesaque}
}

APA:

Mesaque, V. (2024). LLAMA 3 fine-tuned with Unsloth and vitormesaque/irisk dataset. Retrieved from https://huggingface.co/vitormesaque

License

This model is licensed under the MIT License.

Contact

For questions or comments, please contact Vitor Mesaque.

vitormesaque
/

i-llama

{}