A lightweight framework for LLM evaluation

LightEval is a lightweight LLM evaluation suite that Hugging Face has been using internally with the recently released LLM data processing library datatrove and LLM training library nanotron.

We're releasing it with the community in the spirit of building in the open.

Note that it is still very much early so don't expect 100% stability ^^' In case of problems or question, feel free to open an issue!

Clone the repo:

git clone https://github.com/huggingface/lighteval.git
cd lighteval

Create a virtual environment using virtualenv or conda depending on your preferences. We require Python 3.10 or above:

conda create -n lighteval python=3.10 && conda activate lighteval

Install the dependencies. For the default installation, you just need:

pip install .

If you want to evaluate models with frameworks like accelerate or peft, you will need to specify the optional dependencies group that fits your use case (accelerate,tgi,optimum,quantization,adapters,nanotron):

pip install '.[optional1,optional2]'

The setup tested most is:

pip install '.[accelerate,quantization,adapters]'

If you want to push your results to the Hugging Face Hub, don't forget to add your access token to the environment variable HUGGING_FACE_HUB_TOKEN. You can do this by running:

huggingface-cli login

and pasting your access token.

• to load and push big models/datasets, your machine likely needs Git LFS. You can install it with sudo apt-get install git-lfs
• If you want to run bigbench evaluations, install bigbench pip install "bigbench@https://storage.googleapis.com/public_research_data/bigbench/bigbench-0.0.1.tar.gz"

Lastly, if you intend to push to the code base, you'll need to install the precommit hook for styling tests:

pip install .[dev]
pre-commit install

We provide two main entry points to evaluate models:

• run_evals_accelerate.py: evaluate models on CPU or one or more GPUs using 🤗 Accelerate.
• run_evals_nanotron.py: evaluate models in distributed settings using ⚡️ Nanotron.

For most users, we recommend using the 🤗 Accelerate backend - see below for specific commands.

To evaluate a model on one or more GPUs, first create a multi-gpu config by running:

accelerate config

You can then evaluate a model using data parallelism as follows:

accelerate launch --multi_gpu --num_processes=<num_gpus> run_evals_accelerate.py \
    --model_args="pretrained=<path to model on the hub>" \
    --tasks <task parameters> \
    --output_dir output_dir

Here, --tasks refers to either a comma-separated list of supported tasks from the metadata table in the format:

suite|task|num_few_shot|{0 or 1 to automatically reduce `num_few_shot` if prompt is too long}

or a file path like examples/tasks/recommended_set.txt which specifies multiple task configurations. For example, to evaluate GPT-2 on the Truthful QA benchmark run:

accelerate launch --multi_gpu --num_processes=8 run_evals_accelerate.py \
    --model_args "pretrained=gpt2" \
    --tasks "lighteval|truthfulqa:mc|0|0" \
    --override_batch_size 1 \
    --output_dir="./evals/"

Here, --override_batch_size defines the batch size per device, so the effective batch size will be override_batch_size x num_gpus. To evaluate on multiple benchmarks, separate each task configuration with a comma, e.g.

accelerate launch --multi_gpu --num_processes=8 run_evals_accelerate.py \
    --model_args "pretrained=gpt2" \
    --tasks "leaderboard|truthfulqa:mc|0|0,leaderboard|gsm8k|0|0" \
    --override_batch_size 1 \
    --output_dir="./evals/"

See the examples/tasks/recommended_set.txt file for a list of recommended task configurations.

If you want to evaluate a model by spinning up inference endpoints, or use adapter/delta weights, or more complex configuration options, you can load models using a configuration file. This is done as follows:

accelerate launch --multi_gpu --num_processes=<num_gpus> run_evals_accelerate.py \
    --model_config_path="<path to your model configuration>" \
    --tasks <task parameters> \
    --output_dir output_dir

Examples of possible configuration files are provided in examples/model_configs.

To evaluate models larger that ~40B parameters in 16-bit precision, you will need to shard the model across multiple GPUs to fit it in VRAM. You can do this by passing model_parallel=True and adapting --num_processes to be the number of processes to use for data parallel. For example, on a single node of 8 GPUs, you can run:

# PP=2, DP=4 - good for models < 70B params
accelerate launch --multi_gpu --num_processes=4 run_evals_accelerate.py \
    --model_args="pretrained=<path to model on the hub>,model_parallel=True" \
    --tasks <task parameters> \
    --output_dir output_dir

# PP=4, DP=2 - good for huge models >= 70B params
accelerate launch --multi_gpu --num_processes=2 run_evals_accelerate.py \
    --model_args="pretrained=<path to model on the hub>,model_parallel=True" \
    --tasks <task parameters> \
    --output_dir output_dir

To evaluate a model on all the benchmarks of the Open LLM Leaderboard using a single node of 8 GPUs, run:

accelerate launch --multi_gpu --num_processes=8 run_evals_accelerate.py \
    --model_args "pretrained=<model name>" \
    --tasks examples/tasks/open_llm_leaderboard_tasks.txt \
    --override_batch_size 1 \
    --output_dir="./evals/"

You can also use lighteval to evaluate models on CPU, although note this will typically be very slow for large models. To do so, run:

python run_evals_accelerate.py \
    --model_args="pretrained=<path to model on the hub>"\
    --tasks <task parameters> \
    --output_dir output_dir

Independently of the default tasks provided in lighteval that you will find in the tasks_table.jsonl file, you can use lighteval to evaluate models on tasks that require special processing (or have been added by the community). These tasks have their own evaluation suites and are defined as follows:

• extended: tasks which have complex pre- or post-processing and are added by the lighteval maintainers. See the extended_tasks folder for examples.
• community: tasks which have been added by the community. See the community_tasks folder for examples.
• custom: tasks which are defined locally and not present in the core library. Use this suite if you want to experiment with designing a special metric or task.

For example, to run an extended task like ifeval, you can run:

python run_evals_accelerate.py \
    --model_args "pretrained=HuggingFaceH4/zephyr-7b-beta" \
    --use_chat_template \ # optional, if you want to run the evaluation with the chat template
    --tasks "extended|ifeval|0|0" \
    --output_dir "./evals"

To run a community or custom task, you can use (note the custom_tasks flag):

python run_evals_accelerate.py \
    --model_args="pretrained=<path to model on the hub>"\
    --tasks <task parameters> \
    --custom_tasks <path to your custom or community task> \
    --output_dir output_dir

For example, to launch lighteval on arabic_mmlu:abstract_algebra for HuggingFaceH4/zephyr-7b-beta, run:

python run_evals_accelerate.py \
    --model_args "pretrained=HuggingFaceH4/zephyr-7b-beta" \
    --use_chat_template \ # optional, if you want to run the evaluation with the chat template
    --tasks "community|arabic_mmlu:abstract_algebra|5|1" \
    --custom_tasks "community_tasks/arabic_evals" \
    --output_dir "./evals"

lighteval was originally built on top of the great Eleuther AI Harness (we use the latter to power the Open LLM Leaderboard). We also took a lot of inspiration from the amazing HELM, notably for metrics.

Through adding more and more logging functionalities, and making it compatible with increasingly different workflows and model codebases (including 3D parallelism) as well as allowing custom evaluation experiments, metrics and benchmarks, we ended up needing to change the code more and more deeply until lighteval became the small standalone library that it is now.

However, we are very grateful to the Harness and HELM teams for their continued work on better evaluations.

lighteval is supposed to be used as a standalone evaluation library.

• To run the evaluations, you can use run_evals_accelerate.py or run_evals_nanotron.py.
• src/lighteval contains the core of the lib itself
  • lighteval contains the core of the library, divided in the following section
    • main_accelerate.py and main_nanotron.py are our entry points to run evaluation
    • logging: Our loggers, to display experiment information and push it to the hub after a run
    • metrics: All the available metrics you can use. They are described in metrics, and divided between sample metrics (applied at the sample level, such as a prediction accuracy) and corpus metrics (applied over the whole corpus). You'll also find available normalisation functions.
    • models: Possible models to use. We cover transformers (base_model), with adapter or delta weights, as well as TGI models locally deployed (it's likely the code here is out of date though), and brrr/nanotron models.
    • tasks: Available tasks. The complete list is in tasks_table.jsonl, and you'll find all the prompts in tasks_prompt_formatting.py. Popular tasks requiring custom logic are exceptionally added in the extended tasks.
• examples/tasks contains a list of available tasks you can launch. We advise using tasks in the recommended_set, as it's possible that some of the other tasks need double checking.
• tests contains our test suite, that we run at each PR to prevent regressions in metrics/prompts/tasks, for a subset of important tasks.

If your new task or metric has requirements, add a specific requirements.txt file with your evaluation.

To add a new task, first either open an issue, to determine whether it will be integrated in the core evaluations of lighteval, in the extended tasks, or in the community tasks, and add its dataset on the hub.

• Core evaluations are evaluation which only require standard logic in their metrics and processing, and that we will add to our test suite to ensure non regression through time. They already see a high usage in the community.
• Extended evaluations are evaluations which require custom logic in their metrics (complex normalisation, an LLM as a judge, ...), that we added to facilitate the life of users. They already see a high usage in the community.
• Community evaluations are submissions by the community of new tasks.

A popular community evaluation can move to becoming an extended or core evaluation through time.

Prompt function: find a suitable prompt function in src.lighteval.tasks.task_prompt_formatting.py, or code your own. This function must output a Doc object, which should contain query, your prompt, and either gold, the gold output, or choices and gold_index, the list of choices and index or indices of correct answers. If your query contains an instruction which should not be repeated in a few shot setup, add it to an instruction field.

Summary: create a line summary of your evaluation, in src/lighteval/tasks/tasks_table.jsonl. This summary should contain the following fields:

• name (str), your evaluation name
• suite (list), the suite(s) to which your evaluation should belong. This field allows us to compare different tasks implementation, and is used a task selection to differentiate the versions to launch. At the moment, you'll find the keywords ["helm", "bigbench", "original", "lighteval", "community", "custom"]; for core evals, please choose lighteval.
• prompt_function (str), the name of the prompt function you defined in the step above
• hf_repo (str), the path to your evaluation dataset on the hub
• hf_subset (str), the specific subset you want to use for your evaluation (note: when the dataset has no subset, fill this field with "default", not with None or "")
• hf_avail_splits (list), all the splits available for your dataset (train, valid or validation, test, other...)
• evaluation_splits (list), the splits you want to use for evaluation
• few_shots_split (str, can be null), the specific split from which you want to select samples for your few-shot examples. It should be different from the sets included in evaluation_splits
• few_shots_select (str, can be null), the method that you will use to select items for your few-shot examples. Can be null, or one of:
  • balanced selects examples from the few_shots_split with balanced labels, to avoid skewing the few shot examples (hence the model generations) towards one specific label
  • random selects examples at random from the few_shots_split
  • random_sampling selects new examples at random from the few_shots_split for every new item, but if a sampled item is equal to the current one, it is removed from the available samples
  • random_sampling_from_train selects new examples at random from the few_shots_split for every new item, but if a sampled item is equal to the current one, it is kept! Only use this if you know what you are doing.
  • sequential selects the first n examples of the few_shots_split
• generation_size (int), the maximum number of tokens allowed for a generative evaluation. If your evaluation is a log likelihood evaluation (multi-choice), this value should be -1
• stop_sequence (list), a list of strings acting as end of sentence tokens for your generation
• metric (list), the metrics you want to use for your evaluation (see next section for a detailed explanation)
• output_regex (str), A regex string that will be used to filter your generation. (Genrative metrics will only select tokens that are between the first and the second sequence matched by the regex. For example, for a regex matching \n and a generation \nModel generation output\nSome other text the metric will only be fed with Model generation output)
• frozen (bool), for now is set to False, but we will steadily pass all stable tasks to True.
• trust_dataset (bool), set to True if you trust the dataset.

Make sure you can launch your model with your new task using --tasks lighteval|yournewtask|2|0.

Copy the community_tasks/_template.yml to community_tasks/yourevalname.py and edit it to add your custom tasks (the parameters you can use are explained above). It contains an interesting mechanism if the dataset you are adding contains a lot of subsets.

Make sure you can launch your model with your new task using --tasks community|yournewtask|2|0 --custom_tasks community_tasks/yourevalname.py.

First check if you can use one of the parametrized functions in src.lighteval.metrics.metrics_corpus or src.lighteval.metrics.metrics_sample.

If not, you can use the custom_task system to register your new metric:

• create a new python file which should contain the full logic of your metric.
• the file also needs to start with these imports

from aenum import extend_enum
from lighteval.metrics import Metrics

# And any other class you might need to redefine your specific metric, depending on whether it's a sample or corpus metric.

• and to end with the following, so that it adds your metric to our metrics list when loaded as a module.

# Adds the metric to the metric list!
extend_enum(Metrics, "metric_name", metric_function)
if __name__ == "__main__":
    print("Imported metric")

You can then give your custom metric to lighteval by using --custom-tasks path_to_your_file when launching it.

To see an example of a custom metric added along with a custom task, look at examples/tasks/custom_tasks_with_custom_metrics/ifeval/ifeval.py.

These metrics use log-likelihood of the different possible targets.

• loglikelihood_acc (Harness): Fraction of instances where the choice with the best logprob was correct - also exists in a faster version for tasks where the possible choices include only one token (loglikelihood_acc_single_token)
• loglikelihood_acc_norm (Harness): Fraction of instances where the choice with the best logprob, normalized by sequence length, was correct - also exists in a faster version for tasks where the possible choices include only one token (loglikelihood_acc_norm_single_token)
• loglikelihood_acc_norm_nospace (Harness): Fraction of instances where the choice with the best logprob, normalized by sequence length, was correct, with the first space ignored
• loglikelihood_f1 (Harness): Corpus level F1 score of the multichoice selection - also exists in a faster version for tasks where the possible choices include only one token (loglikelihood_f1_single_token)
• mcc (Harness): Matthew's correlation coefficient (measure of agreement between statistical distributions),
• recall_at_1 (Harness): Fraction of instances where the choice with the best logprob was correct - also exists in a faster version for tasks where the possible choices include only one token per choice (recall_at_1_single_token)
• recall_at_2 (Harness): Fraction of instances where the choice with the 2nd best logprob or better was correct - also exists in a faster version for tasks where the possible choices include only one token per choice (recall_at_2_single_token)
• mrr (Harness): Mean reciprocal rank, measure of the quality of a ranking of choices ordered by correctness/relevance - also exists in a faster version for tasks where the possible choices include only one token (mrr_single_token)
• target_perplexity (Harness): Perplexity of the different choices available.
• acc_golds_likelihood: (Harness): A bit different, it actually checks if the average logprob of a single target is above or below 0.5
• multi_f1_numeric: Loglikelihood F1 score for multiple gold targets

All these metrics also exist in a "single token" version (loglikelihood_acc_single_token, loglikelihood_acc_norm_single_token, loglikelihood_f1_single_token, mcc_single_token, recall@2_single_token and mrr_single_token). When the multichoice option compare only one token (ex: "A" vs "B" vs "C" vs "D", or "yes" vs "no"), using these metrics in the single token version will divide the time spent by the number of choices. Single token evals also include:

• multi_f1_numeric (Harness, for CB): computes the f1 score of all possible choices and averages it.

These metrics use log-likelihood of prompt.

• word_perplexity (Harness): Perplexity (log probability of the input) weighted by the number of words of the sequence.
• byte_perplexity (Harness): Perplexity (log probability of the input) weighted by the number of bytes of the sequence.
• bits_per_byte (HELM): Average number of bits per byte according to model probabilities.
• log_prob (HELM): Predicted output's average log probability (input's log prob for language modeling).

These metrics need the model to generate an output. They are therefore slower.

• Base:
  • perfect_exact_match (Harness): Fraction of instances where the prediction matches the gold exactly.
  • exact_match (HELM): Fraction of instances where the prediction matches the gold at the exception of the border whitespaces (= after a strip has been applied to both).
  • quasi_exact_match (HELM): Fraction of instances where the normalized prediction matches the normalized gold (normalization done on whitespace, articles, capitalization, ...). Other variations exist, with other normalizers, such as quasi_exact_match_triviaqa, which only normalizes the predictions after applying a strip to all sentences.
  • prefix_exact_match (HELM): Fraction of instances where the beginning of the prediction matches the gold at the exception of the border whitespaces (= after a strip has been applied to both).
  • prefix_quasi_exact_match (HELM): Fraction of instances where the normalized beginning of the prediction matches the normalized gold (normalization done on whitespace, articles, capitalization, ...)
  • exact_match_indicator: Exact match with some preceding context (before an indicator) removed
  • f1_score_quasi (HELM): Average F1 score in terms of word overlap between the model output and gold, with both being normalized first
  • f1_score: Average F1 score in terms of word overlap between the model output and gold without normalisation
  • f1_score_macro: Corpus level macro F1 score
  • f1_score_macro: Corpus level micro F1 score
• Summarization:
  • rouge (Harness): Average ROUGE score (Lin, 2004)
  • rouge1 (HELM): Average ROUGE score (Lin, 2004) based on 1-gram overlap.
  • rouge2 (HELM): Average ROUGE score (Lin, 2004) based on 2-gram overlap.
  • rougeL (HELM): Average ROUGE score (Lin, 2004) based on longest common subsequence overlap.
  • rougeLsum (HELM): Average ROUGE score (Lin, 2004) based on longest common subsequence overlap.
  • rouge_t5 (BigBench): Corpus level ROUGE score for all available ROUGE metrics
  • faithfulness (HELM): Faithfulness scores based on the SummaC method of Laban et al. (2022).
  • extractiveness (HELM): Reports, based on (Grusky et al., 2018)
    • summarization_coverage: Extent to which the model-generated summaries are extractive fragments from the source document,
    • summarization_density: Extent to which the model-generated summaries are extractive summaries based on the source document,
    • summarization_compression: Extent to which the model-generated summaries are compressed relative to the source document.
  • bert_score (HELM): Reports the average BERTScore precision, recall, and f1 score (Zhang et al., 2020) between model generation and gold summary.
• Translation
  • bleu: Corpus level BLEU score (Papineni et al., 2002) - uses the sacrebleu implementation.
  • bleu_1 (HELM): Average sample BLEU score (Papineni et al., 2002) based on 1-gram overlap - uses the nltk implementation.
  • bleu_4 (HELM): Average sample BLEU score (Papineni et al., 2002) based on 4-gram overlap - uses the nltk implementation.
  • chrf (Harness): Character n-gram matches f-score.
  • ter (Harness): Translation edit/error rate.
• Copyright
  • copyright (HELM): Reports:
    • longest_common_prefix_length: average length of longest common prefix between model generation and reference,
    • edit_distance: average Levenshtein edit distance between model generation and reference,
    • edit_similarity: average Levenshtein edit similarity (normalized by length of longer sequence) between model generation and reference.
• Math:
  • quasi_exact_match_math (HELM): Fraction of instances where the normalized prediction matches the normalized gold (normalization done for math, where latex symbols, units, etc are removed)
  • quasi_exact_match_gsm8k (Harness): Fraction of instances where the normalized prediction matches the normalized gold (normalization done for gsm8k, where latex symbols, units, etc are removed)

To keep compatibility with the Harness for some specific tasks, we ported their evaluations more or less as such. They include drop (for the DROP dataset) and truthfulqa_mc_metrics (for TruthfulQA). In general, except for tasks where the dataset has a very different formatting than usual (an other language, programming language, math, ...), we want to use standard implementations of the above metrics. It makes little sense to have 10 different versions of an exact match depending on the task. However, most of the above metrics are parametrizable so that you can change the normalization applied easily for experimental purposes.

These metrics need both the generation and its logprob. They are not working at the moment, as this fn is not in the AI Harness.

• prediction_perplexity (HELM): Measure of the logprob of a given input.

1. Create a config file for accelerate

compute_environment: LOCAL_MACHINE
distributed_type: MULTI_GPU
downcast_bf16: 'no'
gpu_ids: all
machine_rank: 0
main_training_function: main
mixed_precision: 'no'
num_machines: 1
num_processes: 8
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

2. Create a slurm file

#!/bin/bash
#SBATCH --job-name=kirby-one-node
#SBATCH --nodes=1
#SBATCH --exclusive
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=24
#SBATCH --gres=gpu:8
#SBATCH --mem-per-cpu=11G # This is essentially 1.1T / 96
#SBATCH --partition=production-cluster
#SBATCH --mail-type=ALL
#SBATCH [email protected]

set -x -e
export TMPDIR=/scratch

echo "START TIME: $(date)"

# Activate your relevant virtualenv
source <path_to_your_venv>/activate #or conda activate yourenv

cd <path_to_your_lighteval>/lighteval

export CUDA_LAUNCH_BLOCKING=1
srun accelerate launch --multi_gpu --num_processes=8 run_evals_accelerate.py --model_args "pretrained=your model name" --tasks examples/tasks/open_llm_leaderboard_tasks.txt --override_batch_size 1 --save_details --output_dir=your output dir

pip install build
python3 -m build .