README.md

---
language: 
- en
- de
- es
- fr
tags:
- clir
- colbertx
- plaidx
- xlm-roberta-large
datasets:
- ms_marco
- hltcoe/tdist-msmarco-scores
task_categories:
- text-retrieval
- information-retrieval
task_ids:
- passage-retrieval
- cross-language-retrieval
license: mit
---

# ColBERT-X for English-German/Spanish/French MLIR using Multilingual Translate-Distill

## MLIR Model Setting

- Query language: English
- Query length: 32 token max
- Document language: German/Spanish/French
- 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-clef-mtd-mix-entries-mt5xxl-engeng` is trained with KL-Divergence from the `mt5xxl` MonoT5 reranker
[`unicamp-dl/mt5-13b-mmarco-100k`](https://huggingface.co/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`](https://huggingface.co/datasets/hltcoe/tdist-msmarco-scores/blob/main/t53b-monot5-msmarco-engeng.jsonl.gz). 

### 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 repeated `n` times to iterate through all languages. 

## Usage

To properly load ColBERT-X models from Huggingface Hub, please use the following version of PLAID-X. 
```bash
pip install PLAID-X>=0.3.1
```

Following code snippet loads the model through Huggingface API. 
```python
from colbert.modeling.checkpoint import Checkpoint
from colbert.infra import ColBERTConfig

Checkpoint('hltcoe/plaidx-large-clef-mtd-mix-entries-mt5xxl-engeng', colbert_config=ColBERTConfig())
```

For full tutorial, please refer to the [PLAID-X Jupyter Notebook](https://colab.research.google.com/github/hltcoe/clir-tutorial/blob/main/notebooks/clir_tutorial_plaidx.ipynb), 
which is part of the [SIGIR 2023 CLIR Tutorial](https://github.com/hltcoe/clir-tutorial). 

## BibTeX entry and Citation Info

Please cite the following two papers if you use the model. 


```bibtex
@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}
}
```

```bibtex
@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}
}
```