name: matching-nli
description: "Guides semantic matching, paraphrase, similarity, textual
inference, NLI, DBQA relevance, and ranking dataset discovery in
CLUEDatasetSearch."
disable-model-invocation: true
metadata:
disco-role: operating
license: NO_LICENSE
Matching and NLI Dataset Route
Use this sub-skill when the user intent matches one of these signals:
- semantic matching, paraphrase, similarity, entailment, NLI
- question matching, answer relevance, query-title matching, search ranking
- LCQMC, BQ Corpus, AFQMC, CMNLI, ChineseSTS, CHIP, CAIL SCM, SogouE
Read references/dataset-selection.md for pair/triple matching distinctions and references/troubleshooting.md for QA-vs-matching ambiguity.
Fast workflow
- Restate the user's requested task, language, domain, and output need.
- Search the bundled root index before recommending a row:
python ../../scripts/search_dataset_index.py --category text-matching --query LCQMC
python ../../scripts/search_dataset_index.py --category text-matching --query CAIL
python ../../scripts/search_dataset_index.py --query CMNLI
- Compare title, category, description, keywords, provider, license, paper, and note fields.
- Warn that CLUEDatasetSearch is a catalogue: no dataset files are bundled and external access must be verified.
- If the request belongs to another family, route to the sibling sub-skill named below instead of forcing a weak match.
Boundary routes
Output pattern
When answering, include:
- candidate dataset title and category;
- why it fits the task;
- language/domain/scale hints from the catalogue;
- provider and source URL when present in the bundled index;
- license/access caveat, especially if blank or competition-hosted;
- what the user must verify upstream before download, redistribution, or benchmark use.
Do not tell future agents to open or run files from the original repository checkout.