vlm-eval-harness 0.1.1

CLI-first Python harness for evaluating vision-language models (VLMs) and multimodal LLMs.

VLM Eval Harness

CLI-first Python harness for evaluating vision-language models (VLMs) and multimodal LLMs across multiple benchmarks with unified logging and reports.

Run many VLM benchmarks with one command, with typed configs and reproducible JSON/CSV logs.

Status: ✨ Planning / scaffolding — API and benchmark list may evolve.

Why you might care

• You want a single CLI to run many VLM benchmarks without copy‑pasting ad‑hoc scripts.
• You care about reproducible JSON/CSV logs instead of screenshots of leaderboards.
• You want a typed benchmark protocol so new tasks and model adapters plug in cleanly.

10-second example

Run the working slice — the in-repo toy_qa benchmark against the trivial echo model — straight from the CLI:

# 10-second example: run the toy QA benchmark with the echo model
vlm-eval run \
  --benchmark toy_qa \
  --model echo \
  --config configs/toy_qa_echo.yml

This writes records.jsonl, records.csv, and summary.json under runs/toy_qa_echo/. The echo model only echoes the question, so its accuracy is honestly 0.0 — the example proves the pipeline runs end to end, not model quality. See Quickstart for details.

Architecture

The CLI parses flags into a config, which selects a benchmark and a model adapter; the benchmark yields examples, the model adapter answers them, the evaluation step scores the answers, and everything is written to JSON/CSV logs.

Regenerate the diagram with python scripts/generate_architecture_diagram.py.

Goals

• Provide a single CLI to run multiple VLM benchmarks.
• Support both hosted APIs (e.g., OpenAI, Anthropic) and local models.
• Log results in a unified JSON / CSV schema for leaderboard analysis.
• Make it easy to plug in new tasks, prompts, and models.

Planned Features

• vlm-eval run \ with:
  • --benchmark (or --suite) flag
  • --model / --provider (openai, local, etc.)
  • --config YAML for prompts, decoding, and logging
• Built-in adapters for common VLM / MLLM APIs
• Evaluation datasets wired through the awesome-vlm-evaluation list
• Simple Python API on top of the CLI

Repository Layout (planned)

vlm-eval-harness/
  vlm_eval/
    cli.py           # Typer / Click CLI entrypoint
    config.py        # Pydantic / OmegaConf config definitions
    models/          # Model adapters (OpenAI, local, etc.)
    benchmarks/      # Benchmark runners and task definitions
    logging/         # JSON / CSV logging utilities
    evaluation/      # Metric computation, aggregation
  scripts/
    run_benchmark.py # Example invocations
  configs/
    openai-gpt4o.yml
    local-llava.yml
  tests/
  README.md
  pyproject.toml

Getting Started

This repository is currently in the design and scaffolding phase.

Planned steps:

1. Finalize the minimal benchmark set and logging schema.
2. Implement a thin model adapter interface.
3. Add 1–2 reference configs and example runs.
4. Tag an 0.1.0 pre-release once something reproducible and useful exists.

If you are interested in collaborating, feel free to open an issue with suggestions on:

• Benchmarks you would like to see wired in first.
• Model providers you care about.
• Logging / schema constraints you need for your own analysis.

Quickstart (minimal working slice)

A first runnable vertical slice is implemented: a toy_qa in-repo benchmark and a trivial echo model adapter, wired through the CLI with unified JSON / JSONL / CSV logging.

⚠️ This slice exists to prove the pipeline runs end to end. The echo model only echoes the question back, so it answers nothing correctly — its accuracy on toy_qa is honestly 0.0. There are no real benchmark numbers here yet; real benchmarks and model adapters are TODO.

Install

pip install -e .          # installs the `vlm-eval` CLI
# optional: dev/test extras
pip install -e ".[dev]"

Requires Python 3.10+. Runtime deps are intentionally small: typer, pydantic, pyyaml (plus pytest for tests).

Run the toy benchmark

vlm-eval run --benchmark toy_qa --model echo --config configs/toy_qa_echo.yml

CLI flags (--benchmark, --model, --output-dir, --limit) override the YAML config. Helpers:

vlm-eval list        # show registered benchmarks and models
vlm-eval --version

You can also drive it from Python instead of the CLI:

python scripts/run_toy_qa.py

Output files

Each run writes to <output_dir>/<run_name>/ (default runs/toy_qa_echo/):

File	Contents
records.jsonl	One record per example: {id, question, expected_answer, model_answer, correct}
records.csv	The same per-example records as CSV
summary.json	Aggregated metrics {num_examples, num_correct, accuracy} + run metadata

All metrics are computed from actual model outputs — nothing is hard-coded.

Tests

pytest

The tests run the full toy pipeline and check the log files; they need no API keys and no network access.

What's real vs. TODO

• ✅ CLI (vlm-eval run / list), YAML config, model-adapter interface, toy benchmark, exact-match metric, JSON/JSONL/CSV logging, tests.
• ⏳ TODO: real benchmarks, hosted-API and local-model adapters, richer metrics, and true image inputs (the toy data already carries an image placeholder field so image paths can be plugged in later).

Package layout (as implemented)

vlm_eval_harness/
  cli.py             # Typer CLI: `vlm-eval`
  config.py          # Pydantic RunConfig + YAML loader
  core.py            # shared run_evaluation() orchestrator
  models/            # base adapter + echo adapter + registry
  benchmarks/        # toy_qa benchmark + registry
  evaluation/        # exact-match metric + aggregation
  logging/           # JSON / JSONL / CSV logger
scripts/run_toy_qa.py
configs/toy_qa_echo.yml
tests/