Glossary
- Subject model
An AI system under evaluation: a large language model, a chatbot, or a commercial agricultural advisory service. Subject models answer the benchmark's questions and receive the published scores.
- Judge model
A calibrated large language model used to score a subject model's answer against expert-validated ground truth. AI AgriBench uses several judge models and never allows a model to judge itself.
- Ground truth
The expert-validated reference answer for a benchmark question, derived from land-grant extension publications and peer-reviewed research, or from a documented farmer–expert exchange, and reviewed by agronomists.
- Post-cutoff dataset
A subset of question–answer pairs drawn from source documents published after 30 September 2024, used to check whether a model's performance depends on having seen the source material during pretraining. Results on this subset are marked with an asterisk in the leaderboard.
- Accuracy
The factual correctness of an answer against expert-validated ground truth, scored 0 to 100.
- Completeness
Whether an answer covers the material points an expert answer would cover, scored 0 to 100.
- Relevance
Whether an answer addresses the question that was asked, in its context, scored 0 to 100.
- Conciseness
Whether an answer conveys the material points without padding, scored 0 to 100.
- Main track
The AI AgriBench leaderboard covering technical agronomy questions in English, in a US context, built from land-grant extension publications.
- Smallholders track
The AI AgriBench leaderboard covering real questions asked by smallholder farmers, built from farmer–expert question-and-answer pairs contributed by Digital Green, GAIA/IFPRI and Precision Development.
- Topic category
One of the nine filterable subject areas in the main leaderboard: pests, diseases, weeds, nutrition, soils, seed hybrids, horticulture, water and weather.
- Benchmark release
A versioned, dated publication of results, archived at a permanent URL and never edited after publication. Corrections are issued as a new post that links back to the release.
- Extension publication
Practical agronomic guidance published by a land-grant university's extension service — the documents farmers and advisors actually use, and the basis of the main track's ground truth.
- LLM-as-judge
The practice of using a language model to score another model's output against a reference answer and a rubric. AI AgriBench calibrates its judges against human expert scores and prohibits self-judging. Avoid this term in farmer-facing copy; write "how the answers are scored" instead.
- Consortium member
An organisation that has joined AI AgriBench under its membership terms. Membership does not affect how an organisation's models are scored or reported.