LLM Support-Ticket Triage
Llama-3.2-3B fine-tuned with 4-bit QLoRA for structured-JSON ticket classification
Fine-tuned Llama-3.2-3B with 4-bit QLoRA to classify support tickets into structured JSON, reaching 99% eval token accuracy while training only 0.75% of the model’s parameters.
Beyond the headline number, the project emphasizes disciplined experimentation: I diagnosed overfitting from per-epoch evaluation curves and ran a Weights & Biases LoRA rank ablation (r = 4–32) to find the smallest adapter that holds accuracy.
The data pipeline is Pydantic-validated end-to-end — malformed samples are rejected before they reach training — and the whole project ships with a pytest suite for reproducible training and evaluation runs.
Stack: HuggingFace Transformers + PEFT · QLoRA (bitsandbytes) · Weights & Biases · Pydantic · pytest
Source code: github.com/smasoudrezvani/llm-ticket-triage