evalspec

Gold-label evaluation framework for LLM agents. Measure what your model actually does, track it over time, and catch regressions before they ship.

Quick start

pip install evalspec

# Create a dataset
cat > datasets/analytics.yaml <<EOF
questions:
  - id: Q-001
    question: "How many swaps happened last quarter?"
    gold_answer: "tool_call"
    expected_tool: "get_swap_counts"
    expected_behaviour: "answer_with_citation"
    language: EN
EOF

# Run against OpenAI
export OPENAI_API_KEY="sk-..."
evalspec-run --all --provider openai --model gpt-4o --tag baseline-v1

# Record baseline and check for regressions
evalspec-regression --record baselines/gpt4o.json --provider openai --model gpt-4o
evalspec-regression --check baselines/gpt4o.json --provider openai --model gpt-4o

Gold labels

Every question gets a gold_answer that defines correct behavior:

Agents

CLI tools

Model comparison

evalspec-run --all --provider openai --model gpt-4o --tag gpt4o --report runs/gpt4o.json
evalspec-run --all --provider openai --model gpt-4o-mini --tag gpt4o-mini --report runs/gpt4o-mini.json
evalspec-compare runs/gpt4o.json runs/gpt4o-mini.json --html compare.html

CI gate

# .github/workflows/eval.yml
on: [pull_request]
jobs:
  eval:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: pip install evalspec openai pyyaml
      - run: evalspec-regression --check baselines/gpt4o.json -p openai --model gpt-4o
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

Philosophy

• Gold labels, not heuristics: Every question defines what "correct" means (refuse, clarify, or call a specific tool).
• Version-frozen: Reports include dataset hashes so you know exactly which corpus a score refers to.
• Held-out split: 80/20 stratified split prevents overfitting to the eval set.
• CI-gated: 3% regression tolerance means prompt or model changes don't silently degrade quality.

License

MIT