You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

EXAMI Domain Reranker Adapters

Domain-specific LoRA adapters for Qwen/Qwen3-VL-Reranker-2B, covering the 24 EXAMI domains. Built from the spec in RERANKER_TRAINING_PLAN.md (custom pointwise yes/no trainer; LoRA r=32/α=64 attention-only, 1 epoch, lr 2e-4; hard negatives mined with Qwen3-VL-Embedding-2B; synthetic queries via Claude Sonnet 4.6).

Design: 5 adapters + base for the Math/CS/IT cluster

21 of 24 domains are served by 5 adapters. Mathematics, Computer_Science, and Information_Technology are routed to the BASE reranker (no adapter) — see "Math cluster" below.

Adapter	Domains	Hard-eval MRR (base → adapter)
physical-sci-eng	Physics, Chemistry, Engineering, Earth_and_Environmental_Sciences	0.961 → 0.985
life-health	Biology, Medicine, Agriculture_and_Veterinary	0.954 → 0.998
social-behavioral	Psychology, Economics, Social_Sciences, Education	0.973 → 0.988
business-law-prof	Business, Law, Communications_Journalism_and_Information, Services	0.968 → 0.990
humanities-arts	History, Philosophy, Literature, Religion_and_Theology, Art_and_Design, Music	0.890 → 0.985

Biggest per-domain lifts (R@1, base→adapter, 1+99 hard distractors): Religion 0.70→0.93, Music 0.72→0.97, Literature 0.76→0.98, Agriculture 0.85→0.99, Art 0.85→1.00.

Math cluster → base (evidence)

On 1+99 hard-distractor eval:

Mathematics: base R@1 0.96 vs fine-tuned 0.91 — fine-tuning (grouped and dedicated) consistently degrades Math. Math passages are highly interrelated, so embedding-mined "hard negatives" are often actually valid answers; pointwise training then teaches the model to reject relevant passages. Base is the best Math reranker.
Computer_Science: base R@1 0.99 — already at ceiling; an adapter adds nothing.
Information_Technology: base R@1 0.94 — solid on base. (Only domain with mild adapter upside ~0.94→1.0; a dedicated IT adapter could be added if that gain is worth it.)

Usage

from rerank import DomainReranker          # needs rr_common.py alongside
rr = DomainReranker("pipeline_config.json")
results = rr.rerank("Biology", "How do vaccines create immunity?", [doc1, doc2, ...])
# domains mapped to None (Math/CS/IT) automatically use the base reranker.

Scoring = sigmoid(logit_yes - logit_no) on the model's native judge prompt; results pass the two-signal cutoff (floor 0.05 + largest-gap) from the deployment doc.

Files

<adapter>/ — LoRA weights (adapter_model.safetensors, ~49 MB) + tokenizer
pipeline_config.json — base model, 24-domain→adapter map (Math/CS/IT → null = base), eval
rerank.py, rr_common.py — inference + scoring
eval_hard/*.json — full hard-eval metrics per adapter

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support