msmarco_psg_v1.colbertv2

Description

This is a ColBERT v2 index created via pyterrier_colbert. This is the index that is used in the PLAID-PRF paper.

Usage

# Download and load the index from HuggingFace
import pyterrier as pt
index = pt.Artifact.from_hf('pyterrier/msmarco_psg_v1.colbertv2',
  plaid_mode=True, ncells=4,
  centroid_score_threshold=0.4, ndocs=4096)

# TREC-DL 2019 pt.Experiment:
from pyterrier_colbert.ranking import _prf
from pyterrier.measures import *
pt.Experiment(
    [pt.rewrite.tokenise() >> index.end_to_end()],
    pt.get_dataset("msmarco_passage").get_topics("test-2019"),
    pt.get_dataset("msmarco_passage").get_qrels("test-2019"),
    eval_metrics=[nDCG@10]
)

Benchmarks

TREC 2019

	name	nDCG@10
0	ColBERT-v2	0.738293
1	PLAID-PRF	0.769527

Reproduction

import pyterrier as pt
import pyterrier_colbert.indexing
indexr = pyterrier_colbert.indexing.ColbertV2Indexer("./new_msmarco_colbvert2index", checkpoint="colbert-ir/colbertv2.0", index_name="2bits")
indexr.index(pt.get_dataset("irds:msmarco-passage").get_corpus_iter())

Metadata

{
  "type": "dense_index",
  "format": "colbert",
  "model_checkpoint": "colbert-ir/colbertv2.0",
  "nbits": 2
}