Ollama Judge (Local)
Run evaluations entirely locally with no API key required. Ideal for private data or offline environments.
Setup
Install and start Ollama:
# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh
# Pull a model
ollama pull llama3.2
ollama pull mistral
# Ollama runs at http://localhost:11434 by default
Configuration
AgentEvalConfig config = AgentEvalConfig.builder()
.judgeModel(JudgeModels.ollama("llama3.2"))
.build();
Custom endpoint:
JudgeModels.ollama("llama3.2", "http://my-ollama-server:11434")
Recommended Models
| Model | Parameters | Quality Notes |
|---|---|---|
llama3.2 | 3B/8B | Good general-purpose judge |
mistral | 7B | Strong instruction following |
llama3.1:70b | 70B | Best quality for local evals |
qwen2.5:14b | 14B | Good reasoning |
YAML Configuration
agenteval:
judge:
provider: ollama
model: llama3.2
endpoint: http://localhost:11434 # optional
tip
Ollama is the best choice when:
- Your test data contains sensitive/private information
- You want zero-cost evaluations (just compute)
- You need deterministic, reproducible results