FaisalHejary/multiwoz-embedding-classification

MultiWOZ 2.2 Embedding Classification

Classification data for training text embedding models, built from the utterances of MultiWOZ 2.2 (pfb30/multi_woz_v22, config v2.2). Each row pairs an utterance (the query) with its label names (the positives) and the most confusable wrong label names (the negatives), so labels are treated as documents.

Tasks

Task	Labels	Query instruction	Document instruction
intent	10 user intents, e.g. "find restaurant", "book train"	Given a user message to a tourist information assistant, classify what the user wants to do	Represent this user intent for classification
action	13 dialogue actions, e.g. "asks for information, such as a detail or a preference"	Given a message from a conversation between a tourist and an information assistant, classify the actions it performs	none
topic	7 domains, e.g. "restaurant", "hotel"	Given a message from a conversation between a tourist and an information assistant, classify which topics it is about	none
question	"question or request", "statement"	Classify whether the given message is a question or a statement	none

The task of a row is in its task column. Action and topic rows can have several positives. A turn counts as a question when it has a question-type dialogue act (Request, reqmore, Select, OfferBook) or contains "?".

Fields

• task (string): intent, action, topic or question.
• query_inst (string): instruction for the query.
• query (string): the utterance.
• document_inst (string or null): instruction for the label documents.
• positives (list of string): correct label names.
• negatives (list of string): 1 to 4 wrong label names, hardest first.

How it was built

Every split went through the same steps, scored with Qwen/Qwen3-VL-Embedding-2B:

1. Rank filter: drop a row when any positive is not in the model's top 5 labels; keep the 4 negatives the model scores highest.
2. Easy-row downsampling: rows whose margin (weakest positive score minus hardest negative score) is at least 0.10 are kept with probability 0.25.
3. Dedupe: keep one row per task and query, compared lowercased without punctuation.
4. Per-label cap: at most 3,000 rows per label of each task, drawn in a random order fixed by a hash of each row.

Because of steps 1 and 2, validation and test are easier than raw MultiWOZ; report benchmark numbers on the original data.

Rows

Task	Train	Validation	Test
intent	21,328	4,649	4,667
action	22,381	7,642	7,558
topic	15,924	7,354	7,312
question	6,000	6,000	6,000
Total	65,633	25,645	25,537

Usage

from datasets import load_dataset

ds = load_dataset("FaisalHejary/multiwoz-embedding-classification")

License and citation

Released under Apache 2.0, the license of the source dataset. If you use it, cite MultiWOZ 2.2:

@inproceedings{zang2020multiwoz,
  title={MultiWOZ 2.2: A Dialogue Dataset with Additional Annotation Corrections and State Tracking Baselines},
  author={Zang, Xiaoxue and Rastogi, Abhinav and Sunkara, Srinivas and Gupta, Raghav and Zhang, Jianguo and Chen, Jindong},
  booktitle={Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, ACL 2020},
  pages={109--117},
  year={2020}
}