Finetuning multi-vector models involves several components: the model itself, datasets, loss functions, training arguments, evaluators, and the trainer class. I'll have a look at each of these components, accompanied by practical examples of how they can be used for finetuning strong multi-vector models.
Lastly, in the Evaluation：#evaluation section, I'll show you that my finetuned multi-vector-encoder/mLateOn-medical：https://huggingface.co/multi-vector-encoder/mLateOn-medical model, trained in 14.5 hours on a single RTX 3090 alongside this blogpost, easily outperforms every general-purpose retrieval model I could find on my medical retrieval evaluation: dense, sparse, lexical, and multi-vector alike.
If you're interested in finetuning dense embedding models, sparse embedding models, or rerankers instead, then consider reading through my prior Training and Finetuning Embedding Models：https://huggingface.co/blog/train-sentence-transformers, Training and Finetuning Sparse Embedding Models：https://huggingface.co/blog/train-sparse-encoder, and Training and Finetuning Reranker Models：https://huggingface.co/blog/train-reranker blogposts.
This blogpost is about training multi-vector models. If you want to learn how to use them, from loading and encoding to indexing in vector databases, see the companion Multi-Vector (Late Interaction) Embedding Models with Sentence Transformers：https://huggingface.co/blog/multi-vector-encoder blogpost.
A dense embedding model compresses a whole text into a single vector, and similarity is one dot product between two such summaries. A multi-vector model (also called a late-interaction or ColBERT-style model) skips that compression. It keeps one small vector per token and scores a query against a document with the MaxSim operator, where every query token finds its best-matching document token and the scores are summed. Token-level matching preserves exactly the fine-grained signals that a single vector has to average away, which usually means stronger retrieval, at the cost of a bigger index.
The companion Multi-Vector Embedding Models：https://huggingface.co/blog/multi-vector-encoder blogpost covers the architecture, encoding, scoring, and indexing in detail, so I'll keep this section short and get to the training.
Finetuning multi-vector models significantly improves their retrieval performance on your specific domain: the vocabulary, the query style, and the notion of relevance all differ between web search, legal discovery, code search, and scientific literature review. Because queries and documents are matched token by token, multi-vector models pick up fine-grained domain signals that single-vector models tend to average away, and they respond very well to even modest amounts of in-domain finetuning data.
Beyond that, most released retrieval models were configured for short passages. The classic ColBERT checkpoints truncate documents at 180 or 300 tokens, and many popular dense models at 256 or 512, because their MS MARCO-style training data rarely goes beyond that. If your documents are long, these models silently discard most of every document before scoring it. On my medical evaluation with passages averaging 941 tokens, I measured that this truncation costs up to 0.24 NDCG@10, considerably more than any difference between model architectures. When you train your own model, you configure the document length that your data needs.
LightOn ran into this same dynamic with code retrieval, where general LateOn：https://huggingface.co/lightonai/LateOn wasn't enough and they trained LateOn-Code：https://huggingface.co/lightonai/LateOn-Code. Your domain, whether that's medical, legal, financial, or your company's internal documents, is not getting an official model. This blogpost shows you how to build it yourself, in a matter of hours, on a single consumer GPU.
Training MultiVectorEncoder models involves the following components:
Let's take a closer look at each component.
Multi-vector training gives you a real choice of starting point, and it matters more than you might expect.
If you want to further finetune an existing multi-vector model, you don't have to worry about the architecture at all:
The checkpoint brings its own recipe along: its query and document marker tokens, its projection head, its scoring skiplist. For finetuning, you generally want to keep all of that and change only what your data demands. The first thing to check is the length configuration, since many released checkpoints cap documents at 180 to 512 tokens (see Why Finetune?：#why-finetune), and my medical passages run to 1,400 tokens. The mLateOn family already serves the backbone's full 8192 token context, but if your starting checkpoint carries caps, lift them:
With the per-task caps unset, truncation falls back to the tokenizer's model_max_length , which is why I configure that limit at load time above.
I made one more change, adding a punctuation skiplist that excludes punctuation tokens from document-side scoring and storage. In a 4-way ablation (none, punctuation, stopwords, both) it modestly won on quality, and it shrinks the document index by 9.6% on this data for free:
You can also point MultiVectorEncoder at any base transformer, and a fresh, randomly initialized token-level projection is appended for you:
That's the classic ColBERT pipeline: a Transformer producing contextualized token embeddings, a token-level Dense projecting each of them down to 128 dimensions, a MultiVectorMask deciding which tokens count during scoring, and a token-level Normalize . The projection starts random, so training is required before this model is useful. Interestingly, this works with strong dense embedding backbones too. A fresh projection on Alibaba-NLP/gte-modernbert-base：https://huggingface.co/Alibaba-NLP/gte-modernbert-base reached within 0.03 of the existing-checkpoint starting points in my experiments, from nothing but the projection and 25k training pairs.
The classic ColBERT tokenization tricks ( [MASK] query expansion, [Q] / [D] prefix tokens, a document length cap, a punctuation skiplist) are all off by default and configurable. See Creating Custom Models：https://sbert.net/docs/multi_vector_encoder/usage/custom_models.html for the full set. For what it's worth, I tested [MASK] query expansion in four configurations for my domain finetune and none of them made a measurable difference, so don't feel obliged to reach for the classic recipe.
I measured this directly while preparing this blogpost, taking six starting points and training each with the identical recipe on 25k medical question-passage pairs from MIRIAD：https://huggingface.co/datasets/tomaarsen/miriad-4.4M-split, then evaluating on 1,000 held-out questions against a 50,000 passage corpus:
The result surprised me, and it replicated across two model families. *The -unsupervised checkpoints adapt to a new domain far better than their finished siblings, overtaking them despite starting lower. These checkpoints sit after large-scale contrastive pretraining but before supervised finetuning on general retrieval, so they carry all the late-interaction structure with none of the general-purpose tuning that domain training then has to undo. The finished checkpoints, by contrast, barely moved or even regressed, at every learning rate I tried.
So, if the model family you like publishes a pre-supervised checkpoint, start there. If not, a fresh projection on a strong retrieval-pretrained backbone is a close runner-up. Continuing from a fully finished checkpoint is the weakest option for domain adaptation, despite being the most natural-feeling one.
The MultiVectorEncoderTrainer ：https://sbert.net/docs/package_reference/multi_vector_encoder/trainer.html uses datasets.Dataset ：https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset or datasets.DatasetDict ：https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.DatasetDict instances for training and evaluation. You can load data from the Hugging Face Datasets Hub：https://huggingface.co/datasets or use local data in whatever format you prefer (e.g. CSV, JSON, Parquet, Arrow, or SQL).
Note: Lots of public datasets that work out of the box with Sentence Transformers have been tagged with sentence-transformers on the Hugging Face Hub, so you can easily find them on https://huggingface.co/datasets?other=sentence-transformers：https://huggingface.co/datasets?other=sentence-transformers. Consider browsing through these to find ready-to-go datasets that might be useful for your tasks, domains, or languages.
You can use the load_dataset ：https://huggingface.co/docs/datasets/main/en/package_reference/loading_methods#datasets.load_dataset function to load data from datasets on the Hub:
This is the dataset I'll train on in this blogpost: 4.4 million medical questions from MIRIAD：https://huggingface.co/datasets/miriad/miriad-4.4M, each paired with the source passage that contains its answer (averaging 941 tokens). Simple (query, relevant passage) pairs like these are the easiest retrieval training data to collect for your own domain, and as you'll see, they're all you need.
You can also use load_dataset ：https://huggingface.co/docs/datasets/main/en/package_reference/loading_methods#datasets.load_dataset for loading local data in common file formats:
And if your local data requires pre-processing, you can use datasets.Dataset.from_dict ：https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset.from_dict to initialize your dataset with a dictionary of lists:
It is important that your dataset format matches your loss function (or that you choose a loss function that matches your dataset format). Verifying whether a dataset format works with a loss function involves two steps:
There are two multi-vector specific conventions on top of this:
Loss functions quantify how well a model performs for a given batch of data, allowing an optimizer to update the model weights to produce more favourable (i.e., lower) loss values. The right loss function for your task depends on the data you have and what you're trying to achieve. You can find a full list of options in the Loss Overview：https://sbert.net/docs/multi_vector_encoder/loss_overview.html.
For the common case of question-answer or question-passage pairs, the workhorse is in-batch negatives training with MultiVectorMultipleNegativesRankingLoss ：https://sbert.net/docs/package_reference/multi_vector_encoder/losses.html#multivectormultiplenegativesrankingloss, where every other document in the batch acts as a negative for each query. Bigger batches mean more negatives and stronger training, so in practice you'll want its GradCache variant, CachedMultiVectorMultipleNegativesRankingLoss ：https://sbert.net/docs/package_reference/multi_vector_encoder/losses.html#cachedmultivectormultiplenegativesrankingloss, which decouples the effective batch size from what fits on your GPU:
The mini_batch_size parameter bounds the memory by encoding documents in chunks of this size, while the effective contrastive batch size (128 in my run below, and in my ablations bigger batches bought nothing further) stays a free choice. GradCache guarantees identical results regardless of the chunk size, so lower it for smaller GPUs at only a wall-clock cost. When your document lengths vary a lot, consider its sibling mini_batch_num_tokens , which packs each chunk to a total token budget instead of a document count, so a chunk of unusually long documents can never spike your memory (my mini_batch_size=16 at roughly 940 tokens per document corresponds to mini_batch_num_tokens=15_000 ).
One multi-vector specific trap is that the contrastive losses default to scale=1.0 , unlike the dense embedding equivalent which defaults to scale=20.0 . That 20.0 exists because a cosine similarity is a single value in [-1, 1], too narrow a range for a sharp softmax. A MaxSim score instead sums one best-match similarity per query token, so it already spans roughly [0, query_length]: a 32-token query can score up to 32. So don't copy scale=20.0 over from a dense training script, since it would saturate the softmax and kill your gradients.
For distillation from a stronger teacher, which is how the strongest general-purpose late-interaction models are trained, see MultiVectorDistillKLDivLoss ：https://sbert.net/docs/package_reference/multi_vector_encoder/losses.html#multivectordistillkldivloss and the Knowledge Distillation tab in the Training Overview：https://sbert.net/docs/multi_vector_encoder/training_overview.html#trainer documentation.
You can customize the training process using the MultiVectorEncoderTrainingArguments ：https://sbert.net/docs/package_reference/multi_vector_encoder/training_args.html class. This class lets you adjust parameters that can impact training speed and help you understand what's happening during training.
For more information on the most useful training arguments, check out the Multi-Vector Encoder > Training Overview > Training Arguments：https://sbert.net/docs/multi_vector_encoder/training_overview.html#training-arguments. It's worth reading to get the most out of your training.
Here's an example, using the values from my actual training run:
To track your model's performance during training, you can pass an eval_dataset to the trainer for evaluation loss, but concrete retrieval metrics are much more informative. Sentence Transformers includes the following built-in evaluators for multi-vector models:
For domain finetuning, the MultiVectorInformationRetrievalEvaluator ：https://sbert.net/docs/package_reference/multi_vector_encoder/evaluation.html#multivectorinformationretrievalevaluator built from your own held-out data is the one that matters. One tip on constructing it is that the corpus should be hard enough that models can be told apart. In my case the MIRIAD questions are generated from their own source passages, which makes retrieval unusually easy. Against just the 10k gold passages, nearly every model scored above 0.97 NDCG@10. If your evaluation saturates like that, add distractor passages (I use deduplicated passages from the training split) until the scores spread out:
The MultiVectorEncoderTrainer ：https://sbert.net/docs/package_reference/multi_vector_encoder/trainer.html is where all previous components come together. Here is the complete script that trained multi-vector-encoder/mLateOn-medical：https://huggingface.co/multi-vector-encoder/mLateOn-medical, the model from the introduction:
That's the whole recipe: a pre-supervised checkpoint, a million domain pairs, in-batch negatives, full document length, and a higher-than-usual learning rate. The run took 14.5 hours on my single RTX 3090 at a peak of 17.5 GB VRAM, and every one of those choices was the winner of a measured comparison rather than a guess.
For readers on smaller budgets, my scaling experiments put 100k pairs (75 minutes of training) within 0.012 NDCG@10 of the full million-pair run. Most of the gain comes in the first hour.
The MultiVectorEncoder trainer supports various transformers.TrainerCallback ：https://huggingface.co/docs/transformers/main_classes/callback#transformers.TrainerCallback subclasses, including:
Enable these via the report_to training argument, e.g. report_to=["wandb", "codecarbon"] , with the required dependencies installed. It defaults to "none" , and report_to="all" activates every integration whose dependency is installed.
Refer to the Transformers Callbacks documentation：https://huggingface.co/docs/transformers/en/main_classes/callback for more information on these callbacks and how to create your own.
Typically, top-performing general-purpose models are trained on multiple datasets simultaneously. However, this approach can be challenging due to the varying formats of each dataset. Fortunately, the MultiVectorEncoderTrainer ：https://sbert.net/docs/package_reference/multi_vector_encoder/trainer.html allows you to train on multiple datasets without requiring a uniform format. Additionally, it provides the flexibility to apply different loss functions to each dataset. Here are the steps to train with multiple datasets at once:
Each training/evaluation batch will only contain samples from one of the datasets. The order in which batches are sampled from the multiple datasets is defined by the MultiDatasetBatchSamplers ：https://sbert.net/docs/package_reference/sentence_transformer/sampler.html#sentence_transformers.training_args.MultiDatasetBatchSamplers enum, which can be passed to the MultiVectorEncoderTrainingArguments ：https://sbert.net/docs/package_reference/multi_vector_encoder/training_args.html via multi_dataset_batch_sampler . Valid options are:
To find out where the finetuned model stands, I evaluated it against over 50 retrieval model configurations across four architecture families on the MIRIAD evaluation set, built exactly as in the Evaluator：#evaluator section above, with 1,000 held-out medical questions searching 200,000 unique passages (the 10k gold passages hidden among 190k deduplicated distractors from the training split). This corpus is four times the size of the 50,000-passage one from Which starting point should you pick?：#which-starting-point-should-you-pick, so scores are not comparable between the two tables.
The headline results, with the full table in the collapsible below:
The finetuned model tops the table, beating the strongest zero-shot model of any architecture by +0.062 NDCG@10. In other words, the strongest zero-shot model returns the right passage as the very first hit for 75.8% of the queries, while the finetuned model does so for 84.9%, cutting the rank-1 error by more than a third.
The architecture pattern is just as clear, with the top of the table exclusively late interaction. On long documents, one vector per token beats one vector per document, even at matched training and matched backbones. DenseOn and LateOn share training data and architecture except for the head, and the late-interaction sibling wins by +0.12, with the multilingual pair (mDenseOn and mLateOn) replicating this at +0.13. Scale doesn't rescue single vectors either. Qwen3-Embedding-4B：https://huggingface.co/Qwen/Qwen3-Embedding-4B, the strongest dense model with roughly 33x the active (non-embedding) parameters of mine, still stops 0.13 short, and the 8B version scores lower than the 4B.
BM25 also performs surprisingly well, beating every sparse model, every truncation-capped multi-vector model, and all but three dense models: the multi-billion Qwen3-Embedding-4B：https://huggingface.co/Qwen/Qwen3-Embedding-4B and 8B：https://huggingface.co/Qwen/Qwen3-Embedding-8B, and voyage-4-nano：https://huggingface.co/voyageai/voyage-4-nano, which reads its full 32k token context to edge past by just 0.006. Don't expect that to transfer to your own data though. MIRIAD's questions are generated from the passages, so the lexical overlap between a query and its gold passage is far larger than in typical retrieval, and BM25's unlimited context length lets it use every one of those overlapping words while most neural checkpoints truncate. A BM25 baseline is cheap and always worth running, just don't count on this margin.
The full field at a glance, sorted by score and colored by architecture family.
Models marked @N are evaluated with their document length cap lifted to N tokens, since their native caps (180 to 512 tokens) would otherwise truncate the 941-token average passages. For every multi-vector model this lift was worth +0.08 to +0.24 NDCG@10 over the as-served row, and even the dense DenseOn gained +0.03 from the same treatment.
Note that this does not mean that multi-vector-encoder/mLateOn-medical：https://huggingface.co/multi-vector-encoder/mLateOn-medical is the strongest model on all domains. It's simply the strongest in my domain. This is totally fine, as I just need this model to work well on my data.
Don't underestimate the power of finetuning multi-vector models on your domain. Fourteen and a half hours on a single consumer GPU produced a model that no general-purpose retriever comes close to on this data, and the recipe is a single script with no teacher model and no mined negatives!
扫码直达文章详情