lmms-eval

LMMs-Eval: Probing Intelligence in the Real World

We are building the unified evaluation toolkit for frontier models and probing the abilities in real world, shape what we build next.

📚 Documentation | 📖 100+ Tasks | 🌟 30+ Models | ⚡ Quickstart

🏠 Homepage | 💬 Discord | 🤝 Contributing

Why lmms-eval?

Benchmarks decide what gets built next. A model team that trusts its eval numbers can focus on real improvements instead of chasing noise. But the multimodal evaluation ecosystem is fragmented - scattered datasets, inconsistent post-processing, and single-number accuracy scores that hide whether a gain is real or random. Two teams evaluating the same model on the same benchmark routinely report different results.

We believe better evals lead to better models. Good evaluation maps the border of what models can do and shapes what we build next.

We are building lmms-eval and focusing on three core principles:

• Reproducible - One pipeline, deterministic results. Same model, same benchmark, same numbers, every time.
• Efficient - Evaluation should not be the bottleneck, even at large scale. Async serving, adaptive batching, and video I/O optimizations keep your GPUs saturated end to end.
• Trustworthy - Not just accuracy. Confidence intervals, clustered standard errors, paired comparisons, and ongoing research into evaluation methodology. Results you can trust enough to act on.

For how the pipeline works and the concrete mechanisms behind these principles, see How the Evaluation Pipeline Works and Why it’s Efficient and Trustworthy.

What’s New

v0.7 (Feb 2026) - Operational simplicity and pipeline maturity. 25+ new tasks across 8 domains, 2 new model backends, agentic task evaluation (generate_until_agentic), video I/O overhaul with TorchCodec (up to 3.58x faster), safety/red-teaming baselines, efficiency metrics (per-sample token counts, run-level throughput), and streamlined flattened JSONL log output for cleaner post-analysis. Release notes | Changelog.

v0.6 (Feb 2026) - Evaluation as a service. Standalone HTTP eval server, ~7.5x throughput over v0.5, statistically grounded results (CI, paired t-test), 50+ new tasks. Release notes | Changelog.

v0.5 (Oct 2025) - Audio expansion. Comprehensive audio evaluation, response caching, 50+ benchmark variants across audio, vision, and reasoning. Release notes.

Quickstart

Install and run your first evaluation in under 5 minutes:

git clone https://github.com/EvolvingLMMs-Lab/lmms-eval.git
cd lmms-eval && uv pip install -e ".[all]"

# Run a quick evaluation (Qwen2.5-VL on MME, 8 samples)
python -m lmms_eval \
  --model qwen2_5_vl \
  --model_args pretrained=Qwen/Qwen2.5-VL-3B-Instruct \
  --tasks mme \
  --batch_size 1 \
  --limit 8

If it prints metrics, your environment is ready. For the full guide, see docs/getting-started/quickstart.md.

Installation

Using uv (Recommended for consistent environments)

We use uv for package management to ensure all developers use exactly the same package versions. First, install uv:

curl -LsSf https://astral.sh/uv/install.sh | sh

For development with consistent environment:

git clone https://github.com/EvolvingLMMs-Lab/lmms-eval
cd lmms-eval
# Recommend
uv pip install -e ".[all]"
# If you want to use uv sync
# uv sync  # This creates/updates your environment from uv.lock

To run commands:

uv run python -m lmms_eval --help  # Run any command with uv run

To add new dependencies:

uv add <package>  # Updates both pyproject.toml and uv.lock

Alternative Installation

For direct usage from Git:

uv venv eval
uv venv --python 3.12
source eval/bin/activate
# You might need to add and include your own task yaml if using this installation
uv pip install git+https://github.com/EvolvingLMMs-Lab/lmms-eval.git

If you want to test on caption dataset such as coco, refcoco, and nocaps, you will need to have java==1.8.0 to let pycocoeval api to work. If you don’t have it, you can install by using conda

conda install openjdk=8

you can then check your java version by java -version

If you want to test VILA, you should install the following dependencies:

pip install s2wrapper@git+https://github.com/bfshi/scaling_on_scales

Our Development will be continuing on the main branch, and we encourage you to give us feedback on what features are desired and how to improve the library further, or ask questions, either in issues or PRs on GitHub.

Usage Examples

More examples can be found in examples/models

Evaluation with vLLM

Qwen2.5-VL:

bash examples/models/vllm_qwen2vl.sh

Qwen3-VL:

bash examples/models/vllm_qwen3vl.sh

Qwen3.5:

bash examples/models/vllm_qwen35.sh

Evaluation with SGLang

bash examples/models/sglang.sh

Qwen3.5:

bash examples/models/sglang_qwen35.sh

Wan2.2-I2V on VBVR:

bash examples/models/sglang_wan22_vbvr.sh

Evaluation of OpenAI-Compatible Model

bash examples/models/openai_compatible.sh

Evaluation with Ollama (Local)

bash examples/models/ollama.sh

Evaluation of Qwen2.5-VL

bash examples/models/qwen25vl.sh

Evaluation of Qwen3-VL

bash examples/models/qwen3vl.sh

More Parameters

python3 -m lmms_eval --help

Environmental Variables

Before running experiments and evaluations, we recommend you to export following environment variables to your environment. Some are necessary for certain tasks to run.

export OPENAI_API_KEY="<YOUR_API_KEY>"
export HF_HOME="<Path to HF cache>"
export HF_TOKEN="<YOUR_API_KEY>"
export HF_HUB_ENABLE_HF_TRANSFER="1"
export REKA_API_KEY="<YOUR_API_KEY>"
# Other possible environment variables include
# ANTHROPIC_API_KEY,DASHSCOPE_API_KEY etc.

Common Environment Issues

Sometimes you might encounter some common issues for example error related to httpx or protobuf. To solve these issues, you can first try

python3 -m pip install httpx==0.23.3;
python3 -m pip install protobuf==3.20;
# If you are using numpy==2.x, sometimes may cause errors
python3 -m pip install numpy==1.26;
# Sometimes sentencepiece is required for tokenizer to work
python3 -m pip install sentencepiece;

Custom Model Integration

lmms-eval supports two types of models: Chat (recommended) and Simple (legacy).

Chat Models (Recommended) 🌟

• Location: lmms_eval/models/chat/
• Use: doc_to_messages function from task
• Input: Structured ChatMessages with roles (user, system, assistant) and content types (text, image, video, audio)
• Supports: Interleaved multimodal content
• Uses: Model’s apply_chat_template() method
• Reference: lmms_eval/models/chat/qwen2_5_vl.py or lmms_eval/models/chat/qwen3_vl.py

Example input format:

[
    {"role": "user", "content": [
        {"type": "image", "url": <image>},
        {"type": "text", "text": "What's in this image?"}
    ]}
]

Simple Models (Legacy)

• Location: lmms_eval/models/simple/
• Use: doc_to_visual + doc_to_text functions from task
• Input: Plain text with <image> placeholders + separate visual list
• Supports: Limited (mainly images)
• Manual processing: No chat template support
• Reference: lmms_eval/models/simple/instructblip.py

Example input format:

# Separate visual and text
doc_to_visual -> [PIL.Image]
doc_to_text -> "What's in this image?"

Key Differences

Why Use Chat Models?

• ✅ Built-in chat template support
• ✅ Interleaved multimodal content
• ✅ Structured message protocol
• ✅ Better video/audio support
• ✅ Consistent with modern LLM APIs

Chat Model Implementation Example

from lmms_eval.api.registry import register_model
from lmms_eval.api.model import lmms
from lmms_eval.protocol import ChatMessages

@register_model("my_chat_model")
class MyChatModel(lmms):
    is_simple = False  # Use chat interface

    def generate_until(self, requests):
        for request in requests:
            # 5 elements for chat models
            doc_to_messages, gen_kwargs, doc_id, task, split = request.args

            # Get structured messages
            raw_messages = doc_to_messages(self.task_dict[task][split][doc_id])
            messages = ChatMessages(messages=raw_messages)

            # Extract media and apply chat template
            images, videos, audios = messages.extract_media()
            hf_messages = messages.to_hf_messages()
            text = self.processor.apply_chat_template(hf_messages)

            # Generate...

For more details, see the Model Guide.

Custom Dataset Integration

Task Configuration with doc_to_messages

Implement doc_to_messages to transform dataset documents into structured chat messages:

def my_doc_to_messages(doc, lmms_eval_specific_kwargs=None):
    # Extract visuals and text from doc
    visuals = my_doc_to_visual(doc)
    text = my_doc_to_text(doc, lmms_eval_specific_kwargs)

    # Build structured messages
    messages = [{"role": "user", "content": []}]

    # Add visuals first
    for visual in visuals:
        messages[0]["content"].append({"type": "image", "url": visual})

    # Add text
    messages[0]["content"].append({"type": "text", "text": text})

    return messages

YAML Configuration

task: "my_benchmark"
dataset_path: "my-org/my-dataset"
test_split: test
output_type: generate_until

# For chat models (recommended)
doc_to_messages: !function utils.my_doc_to_messages

# OR legacy approach:
doc_to_visual: !function utils.my_doc_to_visual
doc_to_text: !function utils.my_doc_to_text

process_results: !function utils.my_process_results
metric_list:
  - metric: acc

Key Features

doc_to_messages

• Transforms dataset document into structured chat messages
• Returns: List of message dicts with role and content
• Content supports: text, image, video, audio types
• Protocol: Defined in lmms_eval/protocol.py (ChatMessages class)
• Auto-fallback: If not provided, uses doc_to_visual + doc_to_text

For more details, see the Task Guide.

Web UI

LMMS-Eval includes an optional Web UI for interactive evaluation configuration.

Requirements

• Node.js 18+ (for building the frontend, auto-built on first run)

Usage

# Start the Web UI (opens browser automatically)
uv run lmms-eval-ui

# Custom port
LMMS_SERVER_PORT=3000 LMMS_EVAL_TUI_ALLOWED_ORIGINS=http://localhost:3000 uv run lmms-eval-ui

The web UI provides:

• Model selection from all available models
• Task selection with search/filter
• Real-time command preview
• Live evaluation output streaming
• Start/Stop evaluation controls
• Log Viewer for browsing saved evaluation results and samples

For more details, see Web UI README.

HTTP Evaluation Server

LMMS-Eval includes a production-ready HTTP server for remote evaluation workflows.

Why Use Eval Server?

• Decoupled evaluation: Run evaluations on dedicated GPU nodes while training continues
• Async workflow: Submit jobs without blocking training loops
• Queue management: Sequential job processing with automatic resource management
• Remote access: Evaluate models from any machine

Start Server

from lmms_eval.entrypoints import ServerArgs, launch_server

# Configure server
args = ServerArgs(
    host="0.0.0.0",
    port=8000,
    max_completed_jobs=200,
    temp_dir_prefix="lmms_eval_"
)

# Launch server
launch_server(args)

Server runs at http://host:port with auto-generated API docs at /docs

Client Usage

Sync Client:

from lmms_eval.entrypoints import EvalClient

client = EvalClient("http://eval-server:8000")

# Submit evaluation (non-blocking)
job = client.evaluate(
    model="qwen2_5_vl",
    tasks=["mmmu_val", "mme"],
    model_args={"pretrained": "Qwen/Qwen2.5-VL-7B-Instruct"},
    num_fewshot=0,
    batch_size=1,
    device="cuda:0",
)

# Continue training...
# Later, retrieve results
result = client.wait_for_job(job["job_id"])
print(result["result"])

Async Client:

from lmms_eval.entrypoints import AsyncEvalClient

async with AsyncEvalClient("http://eval-server:8000") as client:
    job = await client.evaluate(
        model="qwen3_vl",
        tasks=["mmmu_val"],
        model_args={"pretrained": "Qwen/Qwen3-VL-4B-Instruct"},
    )
    result = await client.wait_for_job(job["job_id"])

Server API Endpoints

Example Workflow

# Training loop pseudocode
for epoch in range(num_epochs):
    train_one_epoch()

    # After every N epochs, evaluate checkpoint
    if epoch % 5 == 0:
        checkpoint_path = f"checkpoints/epoch_{epoch}"

        # Submit async evaluation (non-blocking)
        eval_job = client.evaluate(
            model="vllm",
            model_args={"model": checkpoint_path},
            tasks=["mmmu_val", "mathvista"],
        )

        # Training continues immediately
        print(f"Evaluation job submitted: {eval_job['job_id']}")

# After training completes, retrieve all results
results = []
for job_id in eval_jobs:
    result = client.wait_for_job(job_id)
    results.append(result)

Security Note

⚠️ This server is intended for trusted environments only. Do NOT expose to untrusted networks without additional security layers (authentication, rate limiting, network isolation).

For more details, see the v0.6 release notes.

Frequently Asked Questions

Acknowledgement

lmms_eval is a fork of lm-eval-harness. We recommend you to read through the docs of lm-eval-harness for relevant information.

Below are the changes we made to the original API:

• Build context now only pass in idx and process image and doc during the model responding phase. This is due to the fact that dataset now contains lots of images and we can’t store them in the doc like the original lm-eval-harness otherwise the cpu memory would explode.
• Instance.args (lmms_eval/api/instance.py) now contains a list of images to be inputted to lmms.
• lm-eval-harness supports all HF language models as single model class. Currently this is not possible for lmms because the input/output format of lmms in HF are not yet unified. Therefore, we have to create a new class for each lmms model. This is not ideal and we will try to unify them in the future.

Citations

@misc{zhang2024lmmsevalrealitycheckevaluation,
      title={LMMs-Eval: Reality Check on the Evaluation of Large Multimodal Models},
      author={Kaichen Zhang and Bo Li and Peiyuan Zhang and Fanyi Pu and Joshua Adrian Cahyono and Kairui Hu and Shuai Liu and Yuanhan Zhang and Jingkang Yang and Chunyuan Li and Ziwei Liu},
      year={2024},
      eprint={2407.12772},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2407.12772},
}

@misc{lmms_eval2024,
    title={LMMs-Eval: Accelerating the Development of Large Multimodal Models},
    url={https://github.com/EvolvingLMMs-Lab/lmms-eval},
    author={Bo Li*, Peiyuan Zhang*, Kaichen Zhang*, Fanyi Pu*, Xinrun Du, Yuhao Dong, Haotian Liu, Yuanhan Zhang, Ge Zhang, Chunyuan Li and Ziwei Liu},
    publisher    = {Zenodo},
    version      = {v0.1.0},
    month={March},
    year={2024}
}