EleutherAI/lm-evaluation-harness

• [2026/09] Plugins: model backends, filters, metrics and aggregations can now be registered from your own package instead of forking — declare an lm_eval.* entry point for zero-config discovery, or point --plugins at a local module. See the Plugin Guide.
• [2025/12] CLI refactored with subcommands (run, ls, validate) and YAML config file support via --config. See the CLI Reference and Configuration Guide.
• [2025/12] Lighter install: Base package no longer includes transformers/torch. Install model backends separately: pip install lm_eval[hf], lm_eval[vllm], etc.
• [2025/07] Added think_end_token arg to hf (token/str), vllm and sglang (str) for stripping CoT reasoning traces from models that support it.
• [2025/03] Added support for steering HF models!
• [2025/02] Added SGLang support!
• [2024/09] We are prototyping allowing users of LM Evaluation Harness to create and evaluate on text+image multimodal input, text output tasks, and have just added the hf-multimodal and vllm-vlm model types and mmmu task as a prototype feature. We welcome users to try out this in-progress feature and stress-test it for themselves, and suggest they check out lmms-eval, a wonderful project originally forking off of the lm-evaluation-harness, for a broader range of multimodal tasks, models, and features.
• [2024/07] API model support has been updated and refactored, introducing support for batched and async requests, and making it significantly easier to customize and use for your own purposes. To run Llama 405B, we recommend using VLLM's OpenAI-compliant API to host the model, and use the local-completions model type to evaluate the model.
• [2024/07] New Open LLM Leaderboard tasks have been added ! You can find them under the leaderboard task group.

A new v0.4.0 release of lm-evaluation-harness is available !

New updates and features include:

• New Open LLM Leaderboard tasks have been added ! You can find them under the leaderboard task group.
• Internal refactoring
• Config-based task creation and configuration
• Easier import and sharing of externally-defined task config YAMLs
• Support for Jinja2 prompt design, easy modification of prompts + prompt imports from Promptsource
• More advanced configuration options, including output post-processing, answer extraction, and multiple LM generations per document, configurable fewshot settings, and more
• Speedups and new modeling libraries supported, including: faster data-parallel HF model usage, vLLM support, MPS support with HuggingFace, and more
• Logging and usability changes
• New tasks including CoT BIG-Bench-Hard, Belebele, user-defined task groupings, and more

Please see our updated documentation pages in docs/ for more details.

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, or in the EleutherAI discord!

This project provides a unified framework to test generative language models on a large number of different evaluation tasks.

Features:

• Over 60 standard academic benchmarks for LLMs, with hundreds of subtasks and variants implemented.
• Support for models loaded via transformers (including quantization via GPTQModel and AutoGPTQ), GPT-NeoX, and Megatron-DeepSpeed, with a flexible tokenization-agnostic interface.
• Support for fast and memory-efficient inference with vLLM.
• Support for commercial APIs including OpenAI, and TextSynth.
• Support for evaluation on adapters (e.g. LoRA) supported in HuggingFace's PEFT library.
• Support for local models and benchmarks.
• Evaluation with publicly available prompts ensures reproducibility and comparability between papers.
• Easy support for custom prompts and evaluation metrics.

The Language Model Evaluation Harness is the backend for 🤗 Hugging Face's popular Open LLM Leaderboard, has been used in hundreds of papers, and is used internally by dozens of organizations including NVIDIA, Cohere, BigScience, BigCode, Nous Research, and Mosaic ML.

To install the lm-eval package from the github repository, run:

git clone --depth 1 https://github.com/EleutherAI/lm-evaluation-harness
cd lm-evaluation-harness
pip install -e .

The base installation provides the core evaluation framework. Model backends must be installed separately using optional extras:

For HuggingFace transformers models:

pip install "lm_eval[hf]"

For vLLM inference:

pip install "lm_eval[vllm]"

For API-based models (OpenAI, Anthropic, etc.):

pip install "lm_eval[api]"

Multiple backends can be installed together:

pip install "lm_eval[hf,vllm,api]"

A detailed table of all optional extras is available at the end of this document.

Use lm-eval -h to see available options, or lm-eval run -h for evaluation options.

List available tasks with:

lm-eval ls tasks

To evaluate a model hosted on the HuggingFace Hub (e.g. GPT-J-6B) on hellaswag you can use the following command (this assumes you are using a CUDA-compatible GPU):

lm_eval --model hf \
    --model_args pretrained=EleutherAI/gpt-j-6B \
    --tasks hellaswag \
    --device cuda:0 \
    --batch_size 8

Additional arguments can be provided to the model constructor using the --model_args flag. Most notably, this supports the common practice of using the revisions feature on the Hub to store partially trained checkpoints, or to specify the datatype for running a model:

lm_eval --model hf \
    --model_args pretrained=EleutherAI/pythia-160m,revision=step100000,dtype="float" \
    --tasks lambada_openai,hellaswag \
    --device cuda:0 \
    --batch_size 8

Models that are loaded via both transformers.AutoModelForCausalLM (autoregressive, decoder-only GPT style models) and transformers.AutoModelForSeq2SeqLM (such as encoder-decoder models like T5) in Huggingface are supported.

Batch size selection can be automated by setting the --batch_size flag to auto. This will perform automatic detection of the largest batch size that will fit on your device. On tasks where there is a large difference between the longest and shortest example, it can be helpful to periodically recompute the largest batch size, to gain a further speedup. To do this, append :N to above flag to automatically recompute the largest batch size N times. For example, to recompute the batch size 4 times, the command would be:

lm_eval --model hf \
    --model_args pretrained=EleutherAI/pythia-160m,revision=step100000,dtype="float" \
    --tasks lambada_openai,hellaswag \
    --device cuda:0 \
    --batch_size auto:4

lm-eval supports evaluating models in GGUF format using the Hugging Face (hf) backend. This allows you to use quantized models compatible with transformers, AutoModel, and llama.cpp conversions.

To evaluate a GGUF model, pass the path to the directory containing the model weights, the gguf_file, and optionally a separate tokenizer path using the --model_args flag.

🚨 Important Note:
If no separate tokenizer is provided, Hugging Face will attempt to reconstruct the tokenizer from the GGUF file — this can take hours or even hang indefinitely. Passing a separate tokenizer avoids this issue and can reduce tokenizer loading time from hours to seconds.

✅ Recommended usage:

lm_eval --model hf \
    --model_args pretrained=/path/to/gguf_folder,gguf_file=model-name.gguf,tokenizer=/path/to/tokenizer \
    --tasks hellaswag \
    --device cuda:0 \
    --batch_size 8

The gguf model type evaluates models hosted by a llama.cpp server (llama-server) through its OpenAI-compatible /v1/completions endpoint:

lm_eval --model gguf \
    --model_args base_url=http://127.0.0.1:8080 \
    --tasks hellaswag

• Requires a llama.cpp release from December 2024 or newer, which returns logprobs in the modern OpenAI format (logprobs.content, see llama.cpp#10783). The deprecated legacy format (token_logprobs) is not supported.
• If the server runs in router mode (multiple models), pass the model name or alias: --model_args base_url=http://127.0.0.1:8080,model=my-model-alias.
• Requests are issued concurrently. The default degree of parallelism is auto-detected from the server's slot count (/props → total_slots, i.e. llama-server's --parallel setting); override it with parallel=<N>. loglikelihood requests are additionally pinned to slots via llama.cpp's id_slot parameter, with consecutive requests sharing a context (e.g. the candidate continuations of one multiple-choice question) assigned to the same slot so they reuse its cached prompt prefix. generate_until requests are left unpinned, since the server's dynamic idle-slot assignment load-balances variable-length generations better.
• loglikelihood is implemented via exact teacher forcing: llama.cpp ignores echo and never returns prompt logprobs, so the continuation is tokenized with the server's /tokenize endpoint (context and continuation encoded separately, matching the HF backend), and each continuation token is scored with its natural prefix as a token-id prompt plus a logit_bias that forces the server to sample it. llama.cpp reports the forced token's pre-sampling logprob and an unbiased top_logprobs, yielding loglikelihoods that match the HF backend to within quantization noise. (A naive alternative — forcing the continuation with a GBNF grammar — does not work: grammar-constrained decoding can pick a fragmented tokenization whose logprobs are not the natural-tokenization loglikelihoods, which systematically deflates scores.)
• loglikelihood_rolling (perplexity tasks such as wikitext) is not implemented for this model type.

We support three main ways of using Hugging Face's accelerate 🚀 library for multi-GPU evaluation.

To perform data-parallel evaluation (where each GPU loads a separate full copy of the model), we leverage the accelerate launcher as follows:

accelerate launch -m lm_eval --model hf \
    --tasks lambada_openai,arc_easy \
    --batch_size 16

(or via accelerate launch --no-python lm_eval).

For cases where your model can fit on a single GPU, this allows you to evaluate on K GPUs K times faster than on one.

WARNING: This setup does not work with FSDP model sharding, so in accelerate config FSDP must be disabled, or the NO_SHARD FSDP option must be used.

The second way of using accelerate for multi-GPU evaluation is when your model is too large to fit on a single GPU.

In this setting, run the library outside the accelerate launcher, but passing parallelize=True to --model_args as follows:

lm_eval --model hf \
    --tasks lambada_openai,arc_easy \
    --model_args parallelize=True \
    --batch_size 16

This means that your model's weights will be split across all available GPUs.

For more advanced users or even larger models, we allow for the following arguments when parallelize=True as well:

• device_map_option: How to split model weights across available GPUs. defaults to "auto".
• max_memory_per_gpu: the max GPU memory to use per GPU in loading the model.
• max_cpu_memory: the max amount of CPU memory to use when offloading the model weights to RAM.
• offload_folder: a folder where model weights will be offloaded to disk if needed.

The third option is to use both at the same time. This will allow you to take advantage of both data parallelism and model sharding, and is especially useful for models that are too large to fit on a single GPU.

accelerate launch --multi_gpu --num_processes {nb_of_copies_of_your_model} \
    -m lm_eval --model hf \
    --tasks lambada_openai,arc_easy \
    --model_args parallelize=True \
    --batch_size 16

To learn more about model parallelism and how to use it with the accelerate library, see the accelerate documentation

Warning: We do not natively support multi-node evaluation using the hf model type! Please reference our GPT-NeoX library integration for an example of code in which a custom multi-machine evaluation script is written.

Note: we do not currently support multi-node evaluations natively, and advise using either an externally hosted server to run inference requests against, or creating a custom integration with your distributed framework as is done for the GPT-NeoX library.

For models that support PyTorch's native Tensor Parallelism (via DTensor), you can shard model weights across GPUs without accelerate's device-map by passing tp_plan=auto in --model_args. Launch with torchrun or accelerate launch:

torchrun --nproc-per-node=4 -m lm_eval \
    --model hf \
    --model_args pretrained=google/gemma-4-31B-it,tp_plan=auto \
    --tasks lambada_openai,arc_easy \
    --batch_size 16

Constraints:

• tp_plan and parallelize=True are mutually exclusive — use one or the other.
• The number of key-value heads in the model must be divisible by --nproc-per-node (the TP degree).
• Requires PyTorch >= 2.4 and a transformers version that exposes a TP plan for the model (v4.47+).

To evaluate a Hugging Face transformers model with steering vectors applied, specify the model type as steered and provide the path to either a PyTorch file containing pre-defined steering vectors, or a CSV file that specifies how to derive steering vectors from pretrained sparsify or sae_lens models (you will need to install the corresponding optional dependency for this method).

Specify pre-defined steering vectors:

import torch

steer_config = {
    "layers.3": {
        "steering_vector": torch.randn(1, 768),
        "bias": torch.randn(1, 768),
        "steering_coefficient": 1,
        "action": "add"
    },
}
torch.save(steer_config, "steer_config.pt")

Specify derived steering vectors:

import pandas as pd

pd.DataFrame({
    "loader": ["sparsify"],
    "action": ["add"],
    "sparse_model": ["EleutherAI/sae-pythia-70m-32k"],
    "hookpoint": ["layers.3"],
    "feature_index": [30],
    "steering_coefficient": [10.0],
}).to_csv("steer_config.csv", index=False)

Run the evaluation harness with steering vectors applied:

lm_eval --model steered \
    --model_args pretrained=EleutherAI/pythia-160m,steer_path=steer_config.pt \
    --tasks lambada_openai,hellaswag \
    --device cuda:0 \
    --batch_size 8

NVIDIA NeMo Framework is a generative AI framework built for researchers and pytorch developers working on language models.

To evaluate a nemo model, start by installing NeMo following the documentation. We highly recommended to use the NVIDIA PyTorch or NeMo container, especially if having issues installing Apex or any other dependencies (see latest released containers). Please also install the lm evaluation harness library following the instructions in the Install section.

NeMo models can be obtained through NVIDIA NGC Catalog or in NVIDIA's Hugging Face page. In NVIDIA NeMo Framework there are conversion scripts to convert the hf checkpoints of popular models like llama, falcon, mixtral or mpt to nemo.

Run a nemo model on one GPU:

lm_eval --model nemo_lm \
    --model_args path=<path_to_nemo_model> \
    --tasks hellaswag \
    --batch_size 32

It is recommended to unpack the nemo model to avoid the unpacking inside the docker container - it may overflow disk space. For that you can run:

mkdir MY_MODEL
tar -xvf MY_MODEL.nemo -c MY_MODEL

By default, only one GPU is used. But we do support either data replication or tensor/pipeline parallelism during evaluation, on one node.

1. To enable data replication, set the model_args of devices to the number of data replicas to run. For example, the command to run 8 data replicas over 8 GPUs is:

torchrun --nproc-per-node=8 --no-python lm_eval \
    --model nemo_lm \
    --model_args path=<path_to_nemo_model>,devices=8 \
    --tasks hellaswag \
    --batch_size 32

1. To enable tensor and/or pipeline parallelism, set the model_args of tensor_model_parallel_size and/or pipeline_model_parallel_size. In addition, you also have to set up devices to be equal to the product of tensor_model_parallel_size and/or pipeline_model_parallel_size. For example, the command to use one node of 4 GPUs with tensor parallelism of 2 and pipeline parallelism of 2 is:

torchrun --nproc-per-node=4 --no-python lm_eval \
    --model nemo_lm \
    --model_args path=<path_to_nemo_model>,devices=4,tensor_model_parallel_size=2,pipeline_model_parallel_size=2 \
    --tasks hellaswag \
    --batch_size 32

Note that it is recommended to substitute the python command by torchrun --nproc-per-node=<number of devices> --no-python to facilitate loading the model into the GPUs. This is especially important for large checkpoints loaded into multiple GPUs.

Not supported yet: multi-node evaluation and combinations of data replication with tensor or pipeline parallelism.

Megatron-LM is NVIDIA's large-scale transformer training framework. This backend allows direct evaluation of Megatron-LM checkpoints without conversion.

Requirements:

• Megatron-LM must be installed or accessible via MEGATRON_PATH environment variable
• PyTorch with CUDA support

Setup:

# Set environment variable pointing to Megatron-LM installation
export MEGATRON_PATH=/path/to/Megatron-LM

Basic usage (single GPU):

lm_eval --model megatron_lm \
    --model_args load=/path/to/checkpoint,tokenizer_type=HuggingFaceTokenizer,tokenizer_model=/path/to/tokenizer \
    --tasks hellaswag \
    --batch_size 1

Supported checkpoint formats:

• Standard Megatron checkpoints (model_optim_rng.pt)
• Distributed checkpoints (.distcp format, auto-detected)

The Megatron-LM backend supports the following parallelism modes:

Data Parallelism (4 GPUs, each with full model replica):

torchrun --nproc-per-node=4 -m lm_eval --model megatron_lm \
    --model_args load=/path/to/checkpoint,tokenizer_model=/path/to/tokenizer,devices=4 \
    --tasks hellaswag

Tensor Parallelism (TP=2):

torchrun --nproc-per-node=2 -m lm_eval --model megatron_lm \
    --model_args load=/path/to/checkpoint,tokenizer_model=/path/to/tokenizer,devices=2,tensor_model_parallel_size=2 \
    --tasks hellaswag

Expert Parallelism for MoE models (EP=4):

torchrun --nproc-per-node=4 -m lm_eval --model megatron_lm \
    --model_args load=/path/to/moe_checkpoint,tokenizer_model=/path/to/tokenizer,devices=4,expert_model_parallel_size=4 \
    --tasks hellaswag

Using extra_args for additional Megatron options:

lm_eval --model megatron_lm \
    --model_args load=/path/to/checkpoint,tokenizer_model=/path/to/tokenizer,extra_args="--no-rope-fusion --trust-remote-code" \
    --tasks hellaswag

Pipeline parallelism during evaluation is supported with OpenVINO models

To enable pipeline parallelism, set the model_args of pipeline_parallel. In addition, you also have to set up device to value HETERO:<GPU index1>,<GPU index2> for example HETERO:GPU.1,GPU.0 For example, the command to use pipeline parallelism of 2 is:

lm_eval --model openvino \
    --tasks wikitext \
    --model_args pretrained=<path_to_ov_model>,pipeline_parallel=True \
    --device HETERO:GPU.1,GPU.0

We also support vLLM for faster inference on supported model types, especially faster when splitting a model across multiple GPUs. For single-GPU or multi-GPU — tensor parallel, data parallel, or a combination of both — inference, for example:

lm_eval --model vllm \
    --model_args pretrained={model_name},tensor_parallel_size={GPUs_per_model},dtype=auto,gpu_memory_utilization=0.8,data_parallel_size={model_replicas} \
    --tasks lambada_openai \
    --batch_size auto

To use vllm, do pip install "lm_eval[vllm]". For a full list of supported vLLM configurations, please reference our vLLM integration and the vLLM documentation.

vLLM occasionally differs in output from Huggingface. We treat Huggingface as the reference implementation and provide a script for checking the validity of vllm results against HF.

We support SGLang for efficient offline batch inference. Its Fast Backend Runtime delivers high performance through optimized memory management and parallel processing techniques. Key features include tensor parallelism, continuous batching, and support for various quantization methods (FP8/INT4/AWQ/GPTQ).

To use SGLang as the evaluation backend, please install it in advance via SGLang documents here.

SGLang's server arguments are slightly different from other backends, see here for more information. We provide an example of the usage here:

lm_eval --model sglang \
    --model_args pretrained={model_name},dp_size={data_parallel_size},tp_size={tensor_parallel_size},dtype=auto \
    --tasks gsm8k_cot \
    --batch_size auto

We support ONNX Runtime GenAI for cross-platform evaluation of ONNX LLMs produced by the ONNX Runtime GenAI Model Builder. Unlike the Windows ML backend, this backend runs on Linux, macOS, and Windows and selects execution providers with the cross-platform og.Config API (CPU, CUDA, DirectML, WebGPU, and AMD NPU via VitisAI/RyzenAI).

Install the backend along with the execution-provider wheel that matches your hardware:

# CPU
pip install "lm_eval[onnxruntime-genai]"
# CUDA / DirectML instead of the CPU wheel (mutually exclusive):
#   pip install onnxruntime-genai-cuda
#   pip install onnxruntime-genai-directml

Evaluate a Model Builder ONNX model on a chosen execution provider:

lm_eval --model onnxruntime-genai \
    --model_args pretrained=/path/to/model_builder_output,execution_provider=cuda \
    --tasks hellaswag \
    --batch_size 1

The pretrained path is a Model Builder output directory (containing genai_config.json, the ONNX graph(s), and an HF tokenizer) or a .onnx file inside one. execution_provider defaults to cpu (which runs with no provider appended); pass cuda, dml, VitisAI, etc. for accelerators. Extra provider settings can be passed via provider_options.

We also support running the same Model Builder export through a raw onnxruntime.InferenceSession instead of the GenAI loop. Use this backend when you want scores to come from the exact runtime a deployment uses, or when you need an execution provider that onnxruntime-genai does not build — notably ROCm and MIGraphX for AMD GPUs.

# CPU
pip install "lm_eval[onnxruntime]"
# CUDA / ROCm instead of the CPU wheel (mutually exclusive):
#   pip install onnxruntime-gpu
#   pip install onnxruntime-rocm

lm_eval --model onnxruntime \
    --model_args pretrained=/path/to/model_builder_output,execution_provider=rocm \
    --tasks hellaswag,arc_easy,wikitext \
    --batch_size 1

execution_provider accepts short aliases (cpu, cuda, rocm, migraphx, dml, openvino, tensorrt, vitisai, webgpu, qnn) as well as full ONNX Runtime provider names. As with the onnxruntime-genai backend, leaving it unset honors the providers declared in the export's genai_config.json, so an export built for a specific device runs there without extra flags. Non-CPU providers keep CPUExecutionProvider as a per-node fallback, since Model Builder graphs use contrib ops that not every provider implements.

Because both ONNX backends share one scoring implementation and end up in the same ORT kernels, they agree on the same model and provider; tests/models/test_onnxruntime_parity.py enforces that as a standing check.

We support Windows ML for hardware-accelerated inference on Windows platforms. This enables evaluation on CPU, GPU, and NPU (Neural Processing Unit) devices.

Windows ML? https://learn.microsoft.com/en-us/windows/ai/new-windows-ml/overview

To use Windows ML, install the required dependencies:

pip install wasdk-Microsoft.Windows.AI.MachineLearning[all] wasdk-Microsoft.Windows.ApplicationModel.DynamicDependency.Bootstrap onnxruntime-windowsml onnxruntime-genai-winml

Evaluate an ONNX Runtime GenAI LLM on NPU/GPU/CPU on Windows:

lm_eval --model winml \
    --model_args pretrained=/path/to/onnx/model \
    --tasks mmlu \
    --batch_size 1

Our library also supports the evaluation of models served via several commercial APIs, and we hope to implement support for the most commonly used performant local/self-hosted inference servers.

To call a hosted model, use:

export OPENAI_API_KEY=YOUR_KEY_HERE
lm_eval --model openai-completions \
    --model_args model=davinci-002 \
    --tasks lambada_openai,hellaswag

We also support using your own local inference server with servers that mirror the OpenAI Completions and ChatCompletions APIs.

lm_eval --model local-completions --tasks gsm8k --model_args model=facebook/opt-125m,base_url=http://{yourip}:8000/v1/completions,num_concurrent=1,max_retries=3,tokenized_requests=False,batch_size=16

Note that for externally hosted models, configs such as --device which relate to where to place a local model should not be used and do not function. Just like you can use --model_args to pass arbitrary arguments to the model constructor for local models, you can use it to pass arbitrary arguments to the model API for hosted models. See the documentation of the hosting service for information on what arguments they support.

Models which do not supply logits or logprobs can be used with tasks of type generate_until only, while local models, or APIs that supply logprobs/logits of their prompts, can be run on all task types: generate_until, loglikelihood, loglikelihood_rolling, and multiple_choice.

For more information on the different task output_types and model request types, see our documentation.

A number of other libraries contain scripts for calling the eval harness through their library. These include GPT-NeoX, Megatron-DeepSpeed, and mesh-transformer-jax.

To create your own custom integration you can follow instructions from this tutorial.

If you have a Metal compatible Mac, you can run the eval harness using the MPS back-end by replacing --device cuda:0 with --device mps (requires PyTorch version 2.1 or higher). Note that the PyTorch MPS backend is still in early stages of development, so correctness issues or unsupported operations may exist. If you observe oddities in model performance on the MPS back-end, we recommend first checking that a forward pass of your model on --device cpu and --device mps match.

python write_out.py \
    --tasks <task1,task2,...> \
    --num_fewshot 5 \
    --num_examples 10 \
    --output_base_path /path/to/output/folder

To verify the data integrity of the tasks you're performing in addition to running the tasks themselves, you can use the --check_integrity flag:

lm_eval --model openai \
    --model_args engine=davinci-002 \
    --tasks lambada_openai,hellaswag \
    --check_integrity

For models loaded with the HuggingFace transformers library, any arguments provided via --model_args get passed to the relevant constructor directly. This means that anything you can do with AutoModel can be done with our library. For example, you can pass a local path via pretrained= or use models finetuned with PEFT by taking the call you would run to evaluate the base model and add ,peft=PATH to the model_args argument:

lm_eval --model hf \
    --model_args pretrained=EleutherAI/gpt-j-6b,parallelize=True,load_in_4bit=True,peft=nomic-ai/gpt4all-j-lora \
    --tasks openbookqa,arc_easy,winogrande,hellaswag,arc_challenge,piqa,boolq \
    --device cuda:0

Models provided as delta weights can be easily loaded using the Hugging Face transformers library. Within --model_args, set the delta argument to specify the delta weights, and use the pretrained argument to designate the relative base model to which they will be applied:

lm_eval --model hf \
    --model_args pretrained=Ejafa/llama_7B,delta=lmsys/vicuna-7b-delta-v1.1 \
    --tasks hellaswag

GPTQ quantized models can be loaded using GPTQModel (faster) or AutoGPTQ

GPTQModel: add ,gptqmodel=True to model_args

lm_eval --model hf \
    --model_args pretrained=model-name-or-path,gptqmodel=True \
    --tasks hellaswag

AutoGPTQ: add ,autogptq=True to model_args:

lm_eval --model hf \
    --model_args pretrained=model-name-or-path,autogptq=model.safetensors,gptq_use_triton=True \
    --tasks hellaswag

We support wildcards in task names, for example you can run all of the machine-translated lambada tasks via --task lambada_openai_mt_*.

To save evaluation results provide an --output_path. We also support logging model responses with the --log_samples flag for post-hoc analysis.

To push results and samples to the Hugging Face Hub, first ensure an access token with write access is set in the HF_TOKEN environment variable. Then, use the --hf_hub_log_args flag to specify the organization, repository name, repository visibility, and whether to push results and samples to the Hub - example dataset on the HF Hub. For instance:

lm_eval --model hf \
    --model_args pretrained=model-name-or-path,autogptq=model.safetensors,gptq_use_triton=True \
    --tasks hellaswag \
    --log_samples \
    --output_path results \
    --hf_hub_log_args hub_results_org=EleutherAI,hub_repo_name=lm-eval-results,push_results_to_hub=True,push_samples_to_hub=True,public_repo=False \

This allows you to easily download the results and samples from the Hub, using:

from datasets import load_dataset

load_dataset("EleutherAI/lm-eval-results-private", "hellaswag", "latest")

For a full list of supported arguments, check out the interface guide in our documentation!

You can seamlessly visualize and analyze the results of your evaluation harness runs using both Weights & Biases (W&B) and Zeno.

You can use Zeno to visualize the results of your eval harness runs.

First, head to hub.zenoml.com to create an account and get an API key on your account page. Add this key as an environment variable:

export ZENO_API_KEY=[your api key]

You'll also need to install the lm_eval[zeno] package extra.

To visualize the results, run the eval harness with the log_samples and output_path flags. We expect output_path to contain multiple folders that represent individual model names. You can thus run your evaluation on any number of tasks and models and upload all of the results as projects on Zeno.

lm_eval \
    --model hf \
    --model_args pretrained=EleutherAI/gpt-j-6B \
    --tasks hellaswag \
    --device cuda:0 \
    --batch_size 8 \
    --log_samples \
    --output_path output/gpt-j-6B

Then, you can upload the resulting data using the zeno_visualize script:

python scripts/zeno_visualize.py \
    --data_path output \
    --project_name "Eleuther Project"

This will use all subfolders in data_path as different models and upload all tasks within these model folders to Zeno. If you run the eval harness on multiple tasks, the project_name will be used as a prefix and one project will be created per task.

You can find an example of this workflow in examples/visualize-zeno.ipynb.

With the Weights and Biases integration, you can now spend more time extracting deeper insights into your evaluation results. The integration is designed to streamline the process of logging and visualizing experiment results using the Weights & Biases (W&B) platform.

The integration provide functionalities

• to automatically log the evaluation results,
• log the samples as W&B Tables for easy visualization,
• log the results.json file as an artifact for version control,
• log the <task_name>_eval_samples.json file if the samples are logged,
• generate a comprehensive report for analysis and visualization with all the important metric,
• log task and cli specific configs,
• and more out of the box like the command used to run the evaluation, GPU/CPU counts, timestamp, etc.

First you'll need to install the lm_eval[wandb] package extra. Do pip install lm_eval[wandb].

Authenticate your machine with an your unique W&B token. Visit https://wandb.ai/authorize to get one. Do wandb login in your command line terminal.

Run eval harness as usual with a wandb_args flag. Use this flag to provide arguments for initializing a wandb run (wandb.init) as comma separated string arguments.

lm_eval \
    --model hf \
    --model_args pretrained=microsoft/phi-2,trust_remote_code=True \
    --tasks hellaswag,mmlu_abstract_algebra \
    --device cuda:0 \
    --batch_size 8 \
    --output_path output/phi-2 \
    --limit 10 \
    --wandb_args project=lm-eval-harness-integration \
    --log_samples

In the stdout, you will find the link to the W&B run page as well as link to the generated report. You can find an example of this workflow in examples/visualize-wandb.ipynb, and an example of how to integrate it beyond the CLI.

Check out our open issues and feel free to submit pull requests!

For more information on the library and how everything fits together, see our documentation pages.

To get started with development, first clone the repository and install the dev dependencies:

git clone https://github.com/EleutherAI/lm-evaluation-harness
cd lm-evaluation-harness
pip install -e ".[dev,hf]"

To implement a new task in the eval harness, see this guide.

In general, we follow this priority list for addressing concerns about prompting and other eval details:

1. If there is widespread agreement among people who train LLMs, use the agreed upon procedure.
2. If there is a clear and unambiguous official implementation, use that procedure.
3. If there is widespread agreement among people who evaluate LLMs, use the agreed upon procedure.
4. If there are multiple common implementations but not universal or widespread agreement, use our preferred option among the common implementations. As before, prioritize choosing from among the implementations found in LLM training papers.

These are guidelines and not rules, and can be overruled in special circumstances.

We try to prioritize agreement with the procedures used by other groups to decrease the harm when people inevitably compare runs across different papers despite our discouragement of the practice. Historically, we also prioritized the implementation from Language Models are Few Shot Learners as our original goal was specifically to compare results with that paper.

The best way to get support is to open an issue on this repo or join the EleutherAI Discord server. The #lm-thunderdome channel is dedicated to developing this project and the #release-discussion channel is for receiving support for our releases. If you've used the library and have had a positive (or negative) experience, we'd love to hear from you!

Extras dependencies can be installed via pip install -e ".[NAME]"

These extras install dependencies required to run specific model backends:

These extras install dependencies required for specific evaluation tasks:

@misc{eval-harness,
  author       = {Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and Reynolds, Laria and Schoelkopf, Hailey and Skowron, Aviya and Sutawika, Lintang and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy},
  title        = {The Language Model Evaluation Harness},
  month        = 07,
  year         = 2024,
  publisher    = {Zenodo},
  version      = {v0.4.3},
  doi          = {10.5281/zenodo.12608602},
  url          = {https://zenodo.org/records/12608602}
}