Zero-shot NLI classifiers (for example MoritzLaurer/deberta-v3-large-zeroshot-v2.0) are a common, cheap baseline for intent and yes/no questions, and they are missing from both the providers and the decision benchmark.
What to build
An HFZeroShot provider in src/thinkless/providers/ that uses the zero-shot-classification pipeline for Choice and YesNo questions and returns calibrated probabilities, loaded the same way as HFClassifier (including the Windows download path in _hub.py).
Where to look
src/thinkless/providers/hf.py: HFClassifier shows loading, device selection and the answer format.docs/guides/providers.md: add a section.
Done when
- Unit tests with a stubbed pipeline pass without downloading a model.
- A benchmark run is attached to the pull request (it can go under
results/submitted/).
Zero-shot NLI classifiers (for example
MoritzLaurer/deberta-v3-large-zeroshot-v2.0) are a common, cheap baseline for intent and yes/no questions, and they are missing from both the providers and the decision benchmark.What to build
An
HFZeroShotprovider insrc/thinkless/providers/that uses thezero-shot-classificationpipeline forChoiceandYesNoquestions and returns calibrated probabilities, loaded the same way asHFClassifier(including the Windows download path in_hub.py).Where to look
src/thinkless/providers/hf.py:HFClassifiershows loading, device selection and the answer format.docs/guides/providers.md: add a section.Done when
results/submitted/).