Fine-tuned LLaMA2 7B model on idioms datasets, employing quantization and parameter-efficient fine-tuning (PEFT) techniques to suggest contextually relevant idioms, enhancing communication naturalness and engagement. Explored 4-bit quantization and parameter-efficient fine-tuning (PEFT) to enhance the model's efficiency and performance.
The Idiom Suggestor project involves three main stages:
- LLM_Synth_Data_Gen_Langchain.ipynb:
- Extracting and preparing idiom data from a comprehensive English idiom dictionary.
- Idioms are extracted from a PDF containing an English idiom dictionary using PyPDF2.
- The text is split and processed into manageable chunks, which are then embedded and indexed using OpenAI's embeddings and FAISS for efficient retrieval during training.
- Llama2_Data_Tamplating.ipynb:
- Formatting the data to align with LLaMA's training requirements.
"<s>[INST] {context} [/INST] {response}</s>\"- Data is transformed into LLaMA-compatible formats using pandas and sklearn, ensuring that idioms and their contexts are properly formatted.
- This step is crucial for effective learning during the model training phase.
- Fine_tune_Llama_2.ipynb:
- Training the model using quantization and PEFT to effectively learn idiomatic expressions.
- Fine-tuning is conducted on a pre-trained LLaMA model using advanced techniques such as LoRA and QLoRA for efficient training.
- Specific attention is paid to managing GPU resources, especially in environments with limited hardware capabilities like Google Colab.
- Idioms-Dictionary.pdf - Source dictonary file for idioms.
- final_Idioms_test_Evaluation.xlsx - LLM validation resulsts and Rogue scored
Before you start, ensure you have Python 3.x installed. Then, clone this repository and install the required packages.
git clone https://github.com/yourusername/idiom-suggestor.git
cd idiom-suggestor
# Execute Each cell in Jupyter files- Model: NousResearch/Llama-2-7b-chat-hf
- Quantization: 4-bit precision using BitsAndBytes
- PEFT: LoRA with an attention dimension of 64 and an alpha scaling parameter of 16
- Training Epochs: 5
- Optimizer: AdamW with a cosine learning rate scheduler.
To assess the quality of the generated idiomatic expressions, we use the ROUGE (Recall-Oriented Understudy for Gisting Evaluation) metric.
Ensure you have the rouge-score library installed to use the ROUGE metrics:
pip install rouge-score