Classical NLP tasks (covered by chat LLM)
HF tasks:
text-classification,token-classification,question-answering,zero-shot-classification,table-question-answering,fill-mask. Status: ⛔ not first-class modalities in Locara.
What they are
The pre-LLM era of NLP: BERT-fine-tuned classifiers, NER taggers, extractive QA, NLI-based zero-shot, table-grounded QA, masked LM. Each is a tractable open-weight model and HuggingFace has a rich catalogue, but a modern chat LLM with the right prompt covers all of these tasks well enough that specializing isn’t worth the surface area.
Representative open-weight models (for completeness)
| Task | Model | Params | License | Notes |
|---|---|---|---|---|
| Text classification | DistilBERT, RoBERTa-Large, DeBERTa-V3-Large | 60-300 M | MIT/Apache-2.0 | Fine-tune for any classification task. |
| NER (token classification) | spaCy NER, BERT-base-NER, GLiNER | 110-300 M | MIT | GLiNER is open-vocabulary (zero-shot NER). |
| Extractive QA | DeBERTa-v3-Large-SQuAD | 300 M | MIT | Foundational. |
| Zero-shot classification | bart-large-mnli, DeBERTa-v3-Large-NLI | 400 M | MIT | NLI-based. |
| Table QA | TaPas-large, TAPEX | 110-400 M | Apache-2.0 | Specialized; heavy. |
| Fill-mask | BERT, RoBERTa | 110-340 M | Apache-2.0 | Internal use only. |
Why Locara doesn’t first-class these
A chat LLM (Qwen3-1.7B+) running through
text-to-text:
- Handles every task above with simple prompts.
- Has flexibility a fixed-class classifier doesn’t (zero-shot generalization, multi-task inference, free-form output).
- Costs the same RAM whether it answers a single question or classifies 1000 inputs.
The exception would be apps where:
- Latency: a 100M-param classifier is 30× faster than a 1.5B LLM for batch classification.
- Determinism: extractive QA is deterministic against a passage; LLM Q&A can hallucinate.
- Privacy budget: classifiers don’t see context beyond their input; LLM “remembers” via in-context tokens.
If a reference app emerges where these matter, we can revisit.
What Locara has today
⛔ Nothing dedicated. Implicit support via
text-to-text.
See also
text-to-text— the catch-alltext-ranking— cross-encoder reranker, similar shape but earned its own first-class slot because of the RAG quality multiplierout-of-scope— adjacent- Index:
../modalities-and-models-survey.md