Training and evaluation scripts for the open mDenseOn dense retriever and mLateOn late-interaction retriever. The models cover multilingual, long-context, and code search. We also include the training scripts for their English-only counterparts, DenseOn and LateOn.
For the data recipe, experiments, and results, check out our mDenseOn and mLateOn blog, the DenseOn and LateOn blog and our paper.
The scripts require Python 3.10 or newer and are intended for CUDA GPUs with bfloat16 support.
git clone git@github.com:lightonai/mdenseon-mlateon.git
cd mdenseon-mlateon
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e .You can also install the environment with uv:
uv syncFor faster training and evaluation, install FlashAttention-2 with uv sync --extra flash. The training scripts use it when you add "attn_implementation": "flash_attention_2" to model_kwargs when instantiating the models; the evaluation scripts take a --fa2 flag.
| Stage | Dense | Late interaction |
|---|---|---|
| Multilingual pre-training | scripts/pretrain/multilingual_dense.py |
scripts/pretrain/multilingual_late_interaction.py |
| Multilingual fine-tuning | scripts/finetune/multilingual_dense.py |
scripts/finetune/multilingual_late_interaction.py |
| English pre-training | scripts/pretrain/english_dense.py |
scripts/pretrain/english_late_interaction.py |
| English fine-tuning | scripts/finetune/english_dense.py |
scripts/finetune/english_late_interaction.py |
The multilingual fine-tuning scripts combine a contrastive loss with KL-divergence distillation from stored cross-encoder teacher scores, and cover the multilingual, long-context, and code datasets. The English ones are contrastive-only on the NV-Embed KD dataset.
Every script exposes its options through --help. Start a distributed run with accelerate launch, for example:
accelerate launch scripts/pretrain/multilingual_dense.py --help
accelerate launch scripts/pretrain/multilingual_dense.py \
--model_name jhu-clsp/mmBERT-base \
--batch_size 16384 \
--mini_batch_size 16Outputs are written under output/. Hugging Face datasets and checkpoints are downloaded on first use, so pre-training requires substantial storage as well as multi-GPU compute. Use --stop_at_step for a short trial run before starting a full job.
The scripts/eval/ folder contains multi-GPU MTEB evaluation scripts for both model families. They require uv sync --extra eval.
All scripts pack length-sorted texts into variable-size batches under a character budget (--encode_char_budget, default 3M characters), so short documents form large batches and long documents small ones without per-task batch-size tuning. Encoding OOMs recover automatically: the dense scripts halve the budget and retry the task, re-computing only the results still missing, while the late-interaction script re-encodes the offending batch in two halves. To further speed up evals, pass --fa2 to enable encoding with FlashAttention-2 (see Setup for installing the flash extra). All scripts expose their options through --help.
Dense models can be evaluated in two ways, sharing the same results layout (one subfolder per model; already-completed (task, language) pairs are skipped on rerun, so the two scripts can fill the same results folder):
scripts/eval/dense_sequential.pyruns tasks sequentially, distributing each task's encoding across all GPUs. This is fastest for large, encode-bound tasks (MSMARCO, MLDR, MIRACL, CodeSearchNet).scripts/eval/dense_parallel.pyruns tasks in parallel, each GPU owning its task's encoding end to end. This is fastest for running several small, retrieval-bound tasks (TREC-COVID, FiQA, SciFact, Quora).
python scripts/eval/dense_sequential.py \
--gpus 0,1,2,3,4,5,6,7 --bf16 \
--results_folder results/dense \
--models lightonai/mDenseOn \
--tasks MIRACLRetrievalHardNegatives MultiLongDocRetrieval
python scripts/eval/dense_parallel.py \
--gpus 0,1,2,3,4,5,6,7 --bf16 \
--results_folder results/dense \
--models lightonai/mDenseOn \
--tasks TRECCOVID FiQA2018 SciFact QuoraRetrievalLate-interaction models encode with accelerate across all GPUs and retrieve with FastPLAID index (fast-plaid>=1.6), where tasks run sequentially as the search step is GPU-bound.
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 accelerate launch scripts/eval/late_interaction.py \
--models lightonai/mLateOn \
--tasks MIRACLRetrievalHardNegatives MultiLongDocRetrieval \
--results_folder results/late_interaction- mDenseOn
- mLateOn
- Multilingual collection, including data
- DenseOn
- LateOn
- English-only collection, including data
If you use our code, models or datasets in your research, please consider citing our work:
@misc{sourty2026denseonlateonfullyopen,
title = {DenseOn with the LateOn: Fully Open Dense and Late-Interaction Models for Multilingual, Long-Context, and Code Search},
author = {Raphaël Sourty and Antoine Chaffin and Paulo Roberto Moura Junior and Amélie Chatelain},
year = {2026},
eprint = {2607.27178},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2607.27178},
}This project is licensed under the Apache License 2.0.
