ColBERT-X for English-Chinese/Persian/Russian MLIR using Multilingual Translate-Distill
MLIR Model Setting
- Query language: English
- Query length: 32 token max
- Document language: Chinese/Persian/Russian
- Document length: 180 token max (please use MaxP to aggregate the passage score if needed)
Model Description
Multilingual Translate-Distill is a training technique that produces state-of-the-art MLIR dense retrieval model through translation and distillation.
plaidx-large-neuclir-mtd-round-robin-entries-mt5xxl-engeng
is trained with KL-Divergence from the mt5xxl
MonoT5 reranker
unicamp-dl/mt5-13b-mmarco-100k
inferenced on English MS MARCO training queries and passages.
The teacher scores can be found in
hltcoe/tdist-msmarco-scores
.
Training Parameters
- learning rate: 5e-6
- update steps: 200,000
- nway (number of passages per query): 6 (randomly selected from 50; 2 if using
round-robin-entires
, see below) - per device batch size (number of query-passage set): 8
- training GPU: 8 NVIDIA V100 with 32 GB memory
Mixing Strategies
mix-passages
: languages are randomly assigned to the 6 sampled passages for a given query during training.mix-entries
: all passages in the a given query-passage set are randomly assigned to the same language.round-robin-entires
: for each query, the query-passage set is repeatedn
times to iterate through all languages.
Usage
To properly load ColBERT-X models from Huggingface Hub, please use the following version of PLAID-X.
pip install PLAID-X>=0.3.1
Following code snippet loads the model through Huggingface API.
from colbert.modeling.checkpoint import Checkpoint
from colbert.infra import ColBERTConfig
Checkpoint('hltcoe/plaidx-large-neuclir-mtd-round-robin-entries-mt5xxl-engeng', colbert_config=ColBERTConfig())
For full tutorial, please refer to the PLAID-X Jupyter Notebook, which is part of the SIGIR 2023 CLIR Tutorial.
BibTeX entry and Citation Info
Please cite the following two papers if you use the model.
@inproceedings{mtt,
title = {Neural Approaches to Multilingual Information Retrieval},
author = {Dawn Lawrie and Eugene Yang and Douglas W Oard and James Mayfield},
booktitle = {Proceedings of the 45th European Conference on Information Retrieval (ECIR)},
year = {2023},
doi = {10.1007/978-3-031-28244-7_33},
url = {https://arxiv.org/abs/2209.01335}
}
@inproceedings{mtd,
author = {Eugene Yang and Dawn Lawrie and James Mayfield},
title = {Distillation for Multilingual Information Retrieval},
booktitle = {Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) (Short Paper) (Accepted)},
year = {2024}
url = {https://arxiv.org/abs/2405.00977}
}
- Downloads last month
- 2