TREC Cast 2019

TREC Cast have released a document collection with topics and qrels of which a subset has been annotated such that it is suitable for multi-turn conversational search.

Dataset statistics

Passages: 38,426,252
Topics: 20
Queries: 173

Subsets

CAR + MSMARCO Collection

Together CAR and MSMARCO have a size of 6,13G, so downloading will take a while. You can use the collection as followed:

collection = load_dataset('trec-cast-2019-multi-turn', 'test_collection')

The collection has the following data format:

docno: str
  The document id format is [collection_id_paragraph_id] with collection id and paragraph id separated by an underscore.
  The collection ids are in the set: {MARCO, CAR}. E.g.: CAR_6869dee46ab12f0f7060874f7fc7b1c57d53144a
text: str
  The content of the passage.

Sample

Instead of using the entire data set, you can also download a sample set containing only 200,000 items:

collection = load_dataset('trec-cast-2019-multi-turn', 'test_collection_sample')

Topics

You can get the topics as followed:

topics = load_dataset('trec-cast-2019-multi-turn', 'topics')

The topics have the following dataformat:

qid: str
  Query ID of the format "topicId_questionNumber"
history: str[]
  A list of queries. It can be empty for the first question in a topic.
query: str
  The query

Qrels

You can get the qrels as followed:

qrels = load_dataset('trec-cast-2019-multi-turn', 'qrels')

The qrels have the following data format:

qid: str
  Query ID of the format "topicId_questionNumber"
qrels: List[dict]
  A list of dictionaries with the keys 'docno' and 'relevance'. Relevance is an integer in the range [0, 4]