Language Model Evaluation Harness

Announcement

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

New updates and features include:

• 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!

Overview

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 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.

Install

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

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

We also provide a number of optional dependencies for extended functionality. A detailed table is available at the end of this document.

Basic Usage

User Guide

A user guide detailing the full list of supported arguments is provided here, and on the terminal by calling lm_eval -h. Alternatively, you can use lm-eval instead of lm_eval.

A list of supported tasks (or groupings of tasks) can be viewed with lm-eval --tasks list. Task descriptions and links to corresponding subfolders are provided here.

Hugging Face transformers

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

Note

Just like you can provide a local path to transformers.AutoModel, you can also provide a local path to lm_eval via --model_args pretrained=/path/to/model

Multi-GPU Evaluation with Hugging Face accelerate

We support two 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 of 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.

These two options (accelerate launch and parallelize=True) are mutually exclusive.

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.

NVIDIA nemo models

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

Multi-GPU evaluation with NVIDIA nemo models

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

2. 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.

Tensor + Data Parallel and Optimized Inference with vLLM

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.

Tip

For fastest performance, we recommend using --batch_size auto for vLLM whenever possible, to leverage its continuous batching functionality!

Tip

Passing max_model_len=4096 or some other reasonable default to vLLM through model args may cause speedups or prevent out-of-memory errors when trying to use auto batch size, such as for Mistral-7B-v0.1 which defaults to a maximum length of 32k.

Model APIs and Inference Servers

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 \
    --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-chat-completions --tasks gsm8k --model_args model=facebook/opt-125m,base_url=http://{yourip}:8000/v1

Note that for externally hosted models, configs such as --device and --batch_size 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.

API or Inference Server	Implemented?	--model <xxx> name	Models supported:	Request Types:
OpenAI Completions	✔️	openai-completions, local-completions	All OpenAI Completions API models	generate_until, loglikelihood, loglikelihood_rolling
OpenAI ChatCompletions	✔️	openai-chat-completions, local-chat-completions	All ChatCompletions API models	generate_until (no logprobs)
Anthropic	✔️	anthropic	Supported Anthropic Engines	generate_until (no logprobs)
Anthropic Chat	✔️	anthropic-chat, anthropic-chat-completions	Supported Anthropic Engines	generate_until (no logprobs)
Textsynth	✔️	textsynth	All supported engines	generate_until, loglikelihood, loglikelihood_rolling
Cohere	⌛ - blocked on Cohere API bug	N/A	All cohere.generate() engines	generate_until, loglikelihood, loglikelihood_rolling
Llama.cpp (via llama-cpp-python)	✔️	gguf, ggml	All models supported by llama.cpp	generate_until, loglikelihood, (perplexity evaluation not yet implemented)
vLLM	✔️	vllm	Most HF Causal Language Models	generate_until, loglikelihood, loglikelihood_rolling
Mamba	✔️	mamba_ssm	Mamba architecture Language Models via the mamba_ssm package	generate_until, loglikelihood, loglikelihood_rolling
Huggingface Optimum (Causal LMs)	✔️	openvino	Any decoder-only AutoModelForCausalLM converted with Huggingface Optimum into OpenVINO™ Intermediate Representation (IR) format	generate_until, loglikelihood, loglikelihood_rolling
Neuron via AWS Inf2 (Causal LMs)	✔️	neuronx	Any decoder-only AutoModelForCausalLM supported to run on huggingface-ami image for inferentia2	generate_until, loglikelihood, loglikelihood_rolling
Neural Magic DeepSparse	✔️	deepsparse	Any LM from SparseZoo or on HF Hub with the "deepsparse" tag	generate_until, loglikelihood
Neural Magic SparseML	✔️	sparseml	Any decoder-only AutoModelForCausalLM from SparseZoo or on HF Hub. Especially useful for models with quantization like zoo:llama2-7b-gsm8k_llama2_pretrain-pruned60_quantized	generate_until, loglikelihood, loglikelihood_rolling
Your local inference server!	✔️	local-completions or local-chat-completions (using openai-chat-completions model type)	Any server address that accepts GET requests using HF models and mirror's OpenAI's Completions or ChatCompletions interface	generate_until

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.

Note

For best performance with closed chat model APIs such as Anthropic Claude 3 and GPT-4, we recommend carefully looking at a few sample outputs using --limit 10 first to confirm answer extraction and scoring on generative tasks is performing as expected. providing system="<some system prompt here>" within --model_args for anthropic-chat-completions, to instruct the model what format to respond in, may be useful.

Other Frameworks

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.

Additional Features

Note

For tasks unsuitable for direct evaluation — either due risks associated with executing untrusted code or complexities in the evaluation process — the --predict_only flag is available to obtain decoded generations for post-hoc evaluation.

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.

Note

You can inspect what the LM inputs look like by running the following command:

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

This will write out one text file for each task.

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 \
    --tasks lambada_openai,hellaswag \
    --check_integrity

Advanced Usage Tips

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 by specifying their file names in ,autogptq=NAME (or ,autogptq=True for default names) in the model_args argument:

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_*.

Saving Results

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

Additionally, one can provide a directory with --use_cache to cache the results of prior runs. This allows you to avoid repeated execution of the same (model, task) pairs for re-scoring.

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!

Visualizing Results

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

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.

Weights and Biases

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.

How to Contribute or Learn More?

For more information on the library and how everything fits together, check out all of our documentation pages! We plan to post a larger roadmap of desired + planned library improvements soon, with more information on how contributors can help.

Implementing new tasks

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.

Support

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!

Optional Extras

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

Name	Use
anthropic	For using Anthropic's models
deepsparse	For running NM's DeepSparse models
dev	For linting PRs and contributions
gptq	For loading models with GPTQ
hf_transfer	For speeding up HF Hub file downloads
ifeval	For running the IFEval task
neuronx	For running on AWS inf2 instances
mamba	For loading Mamba SSM models
math	For running math task answer checking
multilingual	For multilingual tokenizers
openai	For using OpenAI's models
optimum	For running Intel OpenVINO models
promptsource	For using PromptSource prompts
sentencepiece	For using the sentencepiece tokenizer
sparseml	For using NM's SparseML models
testing	For running library test suite
unitxt	For IBM's unitxt dataset tasks
vllm	For loading models with vLLM
zeno	For visualizing results with Zeno
---------------	---------------------------------------
all	Loads all extras (not recommended)

Cite as

@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        = {A framework for few-shot language model evaluation},
  month        = 12,
  year         = 2023,
  publisher    = {Zenodo},
  version      = {v0.4.0},
  doi          = {10.5281/zenodo.10256836},
  url          = {https://zenodo.org/records/10256836}
}