// the find
texttron/tevatron
Tevatron - Unified Document Retrieval Toolkit across Scale, Language, and Modality. Demo in SIGIR 2023, SIGIR 2025.
Tevatron is a training and evaluation toolkit for dense retrieval models — think DPR, but extended to LLM-scale retrievers with LoRA fine-tuning, multi-modal inputs, and both PyTorch and JAX/TPU backends. It's for ML researchers and engineers who need to train or benchmark neural retrieval systems on MS MARCO, BEIR, or custom corpora. Academic in origin (CMU, Waterloo), but used in production-grade IR work.
Dual backend support (PyTorch+DeepSpeed and JAX+TPU) is genuinely useful — the JAX path halves training time on TPU v4 compared to A100 for equivalent throughput. LoRA fine-tuning of 7B+ models is first-class, not bolted on. Self-contained HuggingFace datasets mean you can go from zero to a trained MS MARCO retriever without hunting down preprocessing scripts. The GradCache integration handles in-batch negative scaling cleanly — 512 negatives per query on 4 GPUs without exploding memory.
V2 migration is incomplete — the README says so explicitly, and V1 features living on a separate branch means you may hit dead ends if you need something that wasn't ported yet. The JAX path is marked 'experimental' and lives under a doubly-nested experimental package, which is a real warning sign for anything production-bound. No unified config file or experiment tracking — you're juggling DeepSpeed JSONs and dozens of CLI flags with no obvious way to reproduce a run six months later. Multi-modal support (ColPali, DSE, Qwen-VL) exists but only as standalone examples, not first-class training paths in the main retriever.