Concierge
A general model that has read everything knows your domain no better than it knows your competitor's.
- 1
- GPU to serve what comes out
- 100%
- Dataset and weights owned by you
- 0
- Labelling vendors in the loop
The problem
Domain models are a data problem in a modelling costume. The knowledge is usually written down already, across a few hundred pages, and turning that into something trainable is weeks of work before anyone reaches the interesting part.
What we did
- 01
Crawl a domain and keep its structure, so the dataset inherits the hierarchy the content already had.
- 02
Clean, deduplicate and format into instruction pairs, with whatever was discarded kept and countable.
- 03
Fine-tune a small open model, sized so that serving it is a fixed cost you control.
- 04
Evaluate against held-out material from the same domain, because merely sounding right is the expensive failure.
Where it landed
A domain model in days rather than a quarter, light enough for hardware you already own. The dataset and the weights are yours, which matters the second time you retrain.
Built with
- Python
- PyTorch
- LoRA / PEFT
- Crawlers
- Evaluation harness