Neura Market
Tevatron logo

Tevatron

Free

Dense retrieval training framework (bi-encoder + reranker)

FreeFree tier
Inputs: text, image
Type
Open Source
Visit Website GitHub (38 stars)

About Tevatron

Tevatron is an open-source, unified document retrieval toolkit designed for training and fine-tuning neural retrievers at scale, across different languages and modalities. Developed with both PyTorch and JAX backends, it supports billion-scale LLM neural retriever training on GPUs and TPUs. Key capabilities include parameter-efficient fine-tuning with LoRA, integration with vLLM, DeepSpeed, FlashAttention, and gradient accumulation. Tevatron provides self-contained HuggingFace datasets for multi-modal and multilingual neural retrieval and open-domain QA tasks, and allows direct loading and fine-tuning of state-of-the-art pre-trained models such as BGE-Embedding and Instruct-E5 from HuggingFace. It was presented at SIGIR 2023 and SIGIR 2025.

Key Features

Training billion-scale LLM neural retriever on GPUs and TPUs
Parameter efficient tuning with LoRA
Integration with vLLM, DeepSpeed, FlashAttention, and gradient accumulation
Self-contained HuggingFace datasets for multi-modal and multilingual neural retrieval and open-domain QA
Direct loading and fine-tuning of SOTA pre-trained models (BGE-Embedding, Instruct-E5) from HuggingFace
Supports both PyTorch and JAX implementations

Pros & Cons

Pros
  • Open source and freely available
  • Supports billion-scale neural retriever training
  • Parameter-efficient fine-tuning with LoRA reduces resource requirements
  • Integrates with popular high-performance libraries (DeepSpeed, vLLM, FlashAttention)
  • Easy loading and fine-tuning of state-of-the-art models from HuggingFace
  • Multi-modal and multilingual capabilities built-in
Cons
  • Primarily research-focused, may lack production-ready deployment features
  • Requires significant computational resources for billion-scale training
  • Not a turn-key application; requires programming and ML expertise
  • Documentation may be limited for beginners

Best For

Fine-tuning large language models for dense retrieval tasksOpen-domain question answeringMulti-modal retrieval (text and image)Multilingual document retrievalBuilding retrieval-augmented generation (RAG) pipelinesResearch in neural information retrieval

FAQ

What is Tevatron?
Tevatron is an open-source, unified document retrieval toolkit for training and fine-tuning neural retrievers at scale, across languages and modalities. It supports both PyTorch and JAX backends.
How do I install Tevatron?
Installation requires PyTorch (GPU) or JAX (TPU/GPU). Clone the repository, install dependencies (transformers, datasets, peft, deepspeed, accelerate, faiss, etc.), and run 'pip install -e .' in the Tevatron directory.
What models does Tevatron support?
Tevatron supports direct loading and fine-tuning of state-of-the-art pre-trained models such as BGE-Embedding and Instruct-E5 from HuggingFace.
Can I use Tevatron for multi-modal retrieval?
Yes, Tevatron includes self-contained HuggingFace datasets for multi-modal neural retrieval, and supports query images alongside text.