Open veterinary AI benchmarks

Veterinary answers should be
auditable and continuous.

Two reproducible system evaluations test whether clinical claims carry authoritative support—and whether a product remembers the right pet through real-world changes and interference.

Explore the dataset View methods & code
50 evidence cases 24 memory tasks 100% open score inputs

Track 01 · VetEvidenceBench 2.5

More clinical claims you can actually audit.

Voyage delivered far more risk-weighted, citation-backed content across the same 50 veterinary cases. The advantage survives the prespecified competitor-favourable missing-evidence bound.

Voyage v3.2Leader
0.493
Risk-weighted coverage
Citation precision
0.673
High-risk authoritative
0.458
Claim coverage
0.290
GPT-5.6 Sol
0.079
Risk-weighted coverage
Citation precision
0.507
High-risk authoritative
0.064
Favourable coverage bound
0.122
Claude Opus 5
0.025
Risk-weighted coverage
Citation precision
0.319
High-risk authoritative
0.020
Favourable coverage bound
0.064

Track 02 · PetMemoryBench 2.0

The right memory.
For the right pet.

A multi-pet, high-interference workload probes current state, changed facts, cross-pet isolation, and compound continuity.

Voyage 72/72 100% answer accuracy
24/24 strictly reliable tasks
vs
ChatGPT Free 63/72 87.5% answer accuracy
21/24 strictly reliable tasks
Current state12/12
History & change12/12
Cross-pet isolation12/12
Compound continuity12/12

Under this synthetic workload, Voyage preserved pet-specific current state more reliably than the tested ChatGPT product snapshot. The selected-pet routing design favours Voyage; this is a dated product-system comparison, not a claim about every model or memory setting.

Read the full scorecard →

Run it yourself

Inspect every input.
Reproduce every score.

Frozen prompts, dated baselines, adjudications, schemas, and dependency-free scorers are public. Test your own product against the same contracts.

terminal
git clone https://github.com/Voyage-Pet-AI/voyage-ai-vet-evals
cd voyage-ai-vet-evals
python -m pip install -e .

vet-evals score --track vet-evidence \
  benchmarks/vet-evidence/results/adjudications-v2.5.json

vet-evals score --track pet-memory \
  benchmarks/pet-memory/results/adjudications-v2.0.json
01

Freeze first

Use the published cohort, rubric, and aggregation without tuning.

02

Predict direction

Write the expected score direction before any score-moving change.

03

Report the system

Name the product surface, tools, memory settings, and capture date.