monoT5
The open T5-based document reranker that reframed ranking as text generation — still the reference pointwise reranker years later.
About
monoT5 came out of Jimmy Lin's lab at the University of Waterloo and quietly became the reference implementation of neural reranking. The idea is disarmingly simple: instead of training a classifier to score documents, fine-tune a T5 sequence-to-sequence model on the prompt "Query: ... Document: ... Relevant:" and read the probability of the token "true" versus "false" as the relevance score.
The trick worked so well that monoT5 displaced monoBERT as the go-to research reranker almost overnight, and checkpoints fine-tuned on MS MARCO (castorini/monot5-base-msmarco, monot5-3b-msmarco) became the standard baseline that every new reranker since has had to beat. It ships as plain Hugging Face weights with a permissive research footprint, so you can run it anywhere with a GPU.
It is pointwise — one document at a time — which makes it straightforward but blind to comparisons between candidates, and it is slow compared to modern listwise rerankers. Newer models beat it on BEIR, but as a reproducible baseline it has no equal.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- castorini/pygaggle replication logs (author-reported by the castorini group): monoT5-base MRR@10 ~0.379 on the full MS MARCO passage dev set; ~0.383 on the MS MARCO document dev split
- Independent comparison (E2Rank paper table, 2026, BEIR 8-dataset rerank suite): MonoT5-3B averages 53.93 NDCG@10 vs RankT5-3B 55.40 vs ListT5-3B 55.74 — competitive but superseded by newer rerankers
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: the original paper (Nogueira et al., arXiv 2003.06713, 2020) plus a continuously maintained open implementation at github.com/castorini/pygaggle with public replication logs; HF checkpoints (castorini/monot5-base-msmarco) still in heavy use. Strengths: fully open weights and code, the canonical research baseline for reranking, widely reproduced. Docked for: (1) pointwise design loses to listwise successors on BEIR; (2) maintenance is community-paced rather than actively improved. License: research checkpoints; code under the pygaggle repo license.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Paper (Nogueira et al., arXiv 2003.06713, 2020) plus maintained open implementation at castorini/pygaggle with public replication logs; HF checkpoints castorini/monot5-base-msmarco still in heavy use.
- · Replication logs report MRR@10 ~0.379 on full MS MARCO passage dev; independent E2Rank table has MonoT5-3B at 53.93 avg NDCG@10 vs newer rerankers.
- · Canonical research baseline; pointwise design loses to listwise successors on BEIR; maintenance is community-paced.