A real LoRA run. A visible release decision.
Inspect a completed support-triage experiment: authored synthetic tickets, a pinned open model, measured evaluation, and a saved adapter that reproduces the outputs.
This page replays saved results. It does not run training in your browser or provision a hosted model. All tickets are synthetic; there are no customer results here.
Watch the saved run. Read the product brief.
This recording walks through the completed local experiment. It is a replay of measured results, not live training.
Follow the recorded experiment.
A small dataset with an explicit policy.
Four intents: billing, access, integration, and security. Each ticket maps to a three-field JSON decision. Security tickets should escalate with high priority.
The monthly charge appears twice on our card statement.
{
"intent": "billing",
"next_action": "review_invoice",
"priority": "normal"
}Schema validation and exact-duplicate checks passed. Semantic similarity between splits remains possible.
The outputs, including the failures.
No test examples are hidden. Baseline scores include JSON-format failures and are not a claim that prompting cannot solve this task.
“The monthly charge appears twice on our card statement.”
Base model output
The monthly charge appears twice on our card statement. The reason for this discrepancy is that the card statement is not a valid credit card statement. The card statement is a valid credit card statement that is not a valid credit card statement. To resolve this issue, please provide the card statement that is not valid, and we will be able to identify the discrepancy.
LoRA adapter output
{"intent":"billing","next_action":"review_invoice","priority":"normal"}Matches all expected fields{"intent":"billing","next_action":"review_invoice","priority":"normal"}Run it outside the website.
The source contains the dataset, validation, LoRA training, baseline comparison, artifact checks, and evaluation protocol. It uses Transformers and PEFT; it is not a new training algorithm.
Open the reference implementationpython -m prototype.dataset python -m prototype.run \ --device cpu \ --steps 80 \ --output local-runs/support-triage
Use --device mps on a compatible Apple Silicon Mac. CUDA is an unverified code path. QLoRA is not implemented.
Development evidence, with a narrow scope.
- Authored synthetic data only. No customer data or pilot validation.
- 16 test tickets are too few to establish real-world performance. Semantic similarity between splits remains possible.
- Small model and narrow four-class task; not a general customer support assistant.
- No throughput, latency, cost, privacy-leakage, or production reliability benchmark.
- NVIDIA CUDA, QLoRA, NIM, NeMo, and TensorRT-LLM are unverified unless separately measured.
- Open-source library integration is the prototype implementation, not a novel training algorithm.
- Seed and pinned model revision aid repeatability; exact outputs may vary across hardware and package versions.
From your first dataset.
To your next release.
Bring a use case. Explore a model workflow built for your team.
Request access