Fine-Tuning-Gemma-with-LoRA-for-Customer-Support-Automation

This project fine-tunes the Gemma-7b model using LoRA (Low-Rank Adaptation) on a customer support tweet dataset from Hugging Face to generate automated support replies efficiently.

Objective

Train a memory-efficient LLM to generate accurate, context-aware customer support responses.

Tools & Libraries

• Transformers
• PEFT for LoRA
• Hugging Face Datasets
• PyTorch

Steps

1. Load Dataset & Tokenizer

   • Used mo-customer-support-tweets-945k dataset.
   • Loaded and modified the tokenizer with a [PAD] token.

2. Tokenize Dataset

   • Tokenized both input (customer inquiries) and output (responses).
   • Used padding and truncation.

3. Load & Prepare Gemma Model

   • Loaded model in 8-bit with load_in_8bit=True.
   • Configured LoRA (rank=16, alpha=32, dropout=0.1).
   • Applied LoRA to attention modules (q_proj, v_proj).

4. Training Setup

   • Defined TrainingArguments (batch size, epochs, learning rate, fp16).
   • Used Trainer and DataCollatorForSeq2Seq.

5. Train & Save

   • Trained for 3 epochs with gradient accumulation.
   • Saved both model and tokenizer for inference.