evalkit-bench 0.1.2

Production-grade LLM evaluation and benchmark harness

Evalkit-bench

Production-grade LLM evaluation and benchmark harness

Point it at a YAML test suite → get a live terminal table, a self-contained HTML report, regression tracking, and side-by-side model comparisons.

---

What it does

$ evalkit run evals/factual.yaml --model gpt-4o-mini

  ✓  [1/5]  capital-france
  ✓  [2/5]  capital-japan
  ✓  [3/5]  cell-powerhouse
  ✗  [4/5]  speed-of-light
  ✓  [5/5]  pythagorean

╭──────────────────┬──────────────────────────┬──────────────┬──────────┬────────╮
│ ID               │ Prompt                   │ Output       │ Scorers  │ Result │
├──────────────────┼──────────────────────────┼──────────────┼──────────┼────────┤
│ capital-france   │ What is the capital of…  │ Paris        │ exact ✓  │  PASS  │
│ speed-of-light   │ What is the speed of…    │ 299,792 km/s │ exact ✗  │  FAIL  │
╰──────────────────┴──────────────────────────┴──────────────┴──────────┴────────╯

  ████████████████░░░░  80%  4/5 passed · avg 340ms · 1,200 tokens

  Diff vs previous: 1 regression
  HTML report → .evalkit/runs/20260505T120000_factual-v1_report.html

---

Install

pip install evalkit-bench

Add the embedding scorer (optional, ~1 GB download):

pip install "evalkit-bench[embed]"

Set your API keys — create a .env file in your project:

OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...

---

Quick start

Run a suite:

evalkit run evals/example_factual.yaml

Override model at runtime:

evalkit run evals/example_factual.yaml --model gpt-4o
evalkit run evals/example_factual.yaml --model claude-haiku-4-5-20251001

Compare multiple models side-by-side:

evalkit compare evals/example_factual.yaml \
  --models gpt-4o-mini,gpt-4o,claude-haiku-4-5-20251001

╭──────────────────┬─────────────┬─────────┬──────────────╮
│ Case             │ gpt-4o-mini │  gpt-4o │  haiku-4-5   │
├──────────────────┼─────────────┼─────────┼──────────────┤
│ capital-france   │    PASS     │  PASS   │     PASS     │
│ cell-powerhouse  │    PASS     │  PASS   │     FAIL     │
│ speed-of-light   │    PASS     │  PASS   │     PASS     │
├──────────────────┼─────────────┼─────────┼──────────────┤
│ Pass rate        │    100%     │  100%   │     80%      │
│ Avg latency      │    340ms    │  820ms  │    280ms     │
│ Total tokens     │   1,200     │  2,400  │     950      │
╰──────────────────┴─────────────┴─────────┴──────────────╯

---

Writing a benchmark suite

Suites are plain YAML files:

name: "factual-qa-v1"
description: "Tests basic factual recall"
model: "gpt-4o-mini"             # default model
judge_model: "claude-haiku-4-5-20251001"  # used by llm_judge scorer
temperature: 0.0
max_tokens: 512

cases:
  - id: capital-france
    prompt: "What is the capital of France? Reply with just the city name."
    expected: "Paris"
    scorers: [exact]
    tags: [geography, easy]

  - id: summarise-article
    prompt: "Summarise in 2 sentences: {input}"
    input: "The Eiffel Tower was built between 1887 and 1889..."
    expected: "The Eiffel Tower was built in the late 1880s."
    scorers: [llm_judge, embed]
    rubric: "Does the summary capture the key facts without hallucinating?"
    tags: [summarisation, medium]

All case fields

Field	Required	Description
id	✅	Unique identifier within the suite
prompt	✅	Prompt sent to the model. Supports {input} substitution
expected	—	Reference answer (required by exact, contains, embed)
input	—	Replaces {input} in the prompt
scorers	—	List of scorers (default: [exact])
rubric	—	Evaluation guidance for llm_judge
tags	—	Arbitrary labels for filtering
pattern	—	Regex pattern for exact scorer

---

Scorers

Scorer	Passes when	Best for
exact	Output matches expected (case-insensitive). Exact → 1.0, substring → 0.5	Short deterministic answers
contains	Expected string appears anywhere in output	Keywords, function names
llm_judge	Claude rates the output ≥ 3.5 / 5	Open-ended quality, summarisation
embed	Cosine similarity of sentence embeddings ≥ 0.7	Semantic equivalence

Multiple scorers are AND-ed — a case passes only if every scorer passes.

---

Regression tracking

Every run is saved automatically to .evalkit/runs/. After each run, evalkit diffs against the previous run for the same suite:

Diff vs previous: 2 regressions, 1 improvement

Compare any two runs manually:

evalkit diff .evalkit/runs/20260505T090000_factual-v1.json \
             .evalkit/runs/20260505T100000_factual-v1.json

List all runs:

evalkit list-runs

Inspect a specific run:

evalkit show 20260505T120000

---

CLI reference

evalkit run     <suite.yaml>  [--model MODEL] [--output-dir DIR] [--no-report]
evalkit compare <suite.yaml>  --models MODEL1,MODEL2,...
evalkit diff    <run_a.json>  <run_b.json>
evalkit list-runs             [--runs-dir DIR]
evalkit show    <run_id>      [--runs-dir DIR]

CI-friendly: evalkit run exits with 0 if all cases pass, 1 if any fail.

---

HTML report

Every run generates a self-contained HTML report — no CDN, works offline, sendable as an attachment.

• Pass/fail per case with color coding
• Click any row to expand full prompt, output, and judge reasoning
• Regression and improvement badges when diffing against a previous run
• Overall pass-rate bar and summary stats

---

Roadmap

• Gemini and Ollama (local) providers
• Parallel case execution
• Web UI dashboard for run history
• Custom scorer plugins via entry points
• evalkit init — interactive suite generator

---

Contributing

git clone https://github.com/Arman176001/evalkit
cd evalkit
pip install -e ".[dev]"
python -m pytest tests/ -v

All 44 tests run in under 3 seconds with no API calls required.

---

Made by Arman · MIT License