What you are looking at
Fine-tuning means giving an existing AI model extra practice on examples of one specific task. LoRA does this cheaply: instead of changing all 135 million numbers inside the model, it trains a small add-on (about 2% of the size) and leaves the original untouched.
The job here: read a support ticket and answer only with {"category": ..., "priority": ...}. The original model chats instead of following the format. After training it answers in the right format, usually with the right labels.
- Pick a ticket below (these tickets were never used for training).
- Compare the original model with the fine-tuned model and the expected answer.
- Scroll down for the scores over all test tickets, and the training loss curve (loss falling means the model is learning).
Everything on this page is a recording of real runs on a CPU: it does not call a live model. The tickets are synthetic (generated from templates), so the scores show the method works, not how it would do on real customer text.
1. Choose a test ticket
Original model (no fine-tuning)
Pick a ticket above.
Fine-tuned model (base + LoRA adapter)
Ticket and expected answer
2. Scores on held-out tickets
"Seen style" = new tickets written from the same sentence patterns as training. "Unseen wording" = sentence patterns and urgency phrases the model never saw in training. "Exact match" needs valid JSON with both labels right.
3. Training loss
Loss measures how surprised the model is by the correct answer; lower is better. Thin line: each training step. Thick line: smoothed. Dots: validation loss on tickets that are not trained on.
4. Serving it
The repository ships an OpenAI-compatible server, so any OpenAI client can talk to the fine-tuned model:
llmft serve --adapter artifacts/adapter # runs on CPU
curl localhost:8000/v1/chat/completions -H 'content-type: application/json' \
-d '{"model":"triage-lora","messages":[{"role":"user","content":"..."}]}'
For GPUs there is a vLLM config (not executed in this build), and SageMaker and Vertex AI deployment code that was unit-tested but never run against a real cloud account. See the README for the exact status of each piece.