A historic-district newsletter classifier trained on open-web reviews keeps treating lot as a parking space when staff mean a parcel. How should the app fix that mismatch?
Select an answer to reveal the explanation.
Short Explanation
On the open web, “lot” is parking. In the historic district, “lot” is a parcel. Add a small set of district-labeled notes on top of the pretrained model. A new transformer family, a layout tool, or an NCCL shard will not teach that civic sense.
Full Explanation
Domain language shift is fixed by adding in-domain labeled notes on top of the pretrained model, not by inventing a new transformer. A dedicated layer for one civic word is not an associate fix. A parcel mention is still text, so a multimodal layout tool is the wrong track. NCCL sharding does not teach the model district vocabulary.