OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Developer Tools & SDKs

Trove

Trove is an open-source Python toolkit for training and evaluating transformer-based dense retrievers, with modular components for retrieval experiments, dataset manipulation, evaluation and hard-negative mining.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Trove — OpenViST project profile in AI Developer Tools & SDKs.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Trove when a retrieval experiment needs reusable training and evaluation components, flexible relevance-data handling or hard-negative mining.

Who it is for

  • Researchers experimenting with transformer-based dense retrieval.
  • Developers assembling retriever training and evaluation workflows.
  • Teams working with multiple relevance datasets or hard-negative mining.

Documented uses

  • Train dense retrievers using binary or graduated relevance labels.
  • Combine real, synthetic and mined retrieval data in a dataset workflow.
  • Calculate information-retrieval metrics for a retriever.
  • Mine hard negatives for given queries.
Before you choose

Requirements & limits

The public GitHub repository identifies an Apache-2.0 license and documents installation through pip or directly from the repository.

  • The documented examples require query, corpus and relevance data files for training or evaluation.
  • The toolkit is designed to work with Hugging Face Transformers and describes PEFT and distributed training/inference integration.
  • The README documents multi-GPU and multi-node launches with Accelerate and DeepSpeed configuration for training.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Train a dense retriever with binary relevance labels

Set up and run Trove training using positive and negative qrels from local files.

Before you start

  • Trove and its Hugging Face Transformers integration available in the environment.
  • A model checkpoint, plus queries.jsonl, corpus.jsonl, and train_qrel.tsv files in the paths used by the script.
  • Training arguments, model arguments, and data arguments configured for the task.

Steps

  1. Create RetrievalTrainingArguments, ModelArguments, and DataArguments; configure the model checkpoint and data-processing settings.

  2. Load a BiEncoderRetriever from the model arguments and create positive and negative MaterializedQRelConfig objects for the qrels, corpus, and query files.

  3. Build a BinaryDataset using the positive and negative configurations and the model's query and passage formatting methods.

  4. Load a tokenizer, create a RetrievalCollator, then create a RetrievalTrainer with the model, tokenizer, collator, and dataset; call trainer.train() and trainer.save_model().

Expected outcome

Trove runs the training workflow and saves the model using the configured output directory.

Guide limitations

  • BinaryDataset is for binary relevance labels; use MultiLevelDataset for multiple relevance levels.
  • The documented example filters positives with min_score=1 and negatives with max_score=1; adapt these settings to the qrel labels and dataset.
  • The recipe does not specify hardware requirements or guarantee training performance.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗Guide documentation ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
BatsResearch/trove · Official homepage title · retained in its source language
How we review sourcesSuggest a correction