05 Jun 2025

Evaluation harnesses list and notes

Previously:

more generic: 250317-1847 Current LLM evaluation landscape
250505-1019 Evaluating RAG
250415-1640 Evaluating NLG

EleutherAI lm-eval-harness

Link: EleutherAI/lm-evaluation-harness: A framework for few-shot evaluation of language models.

Running

Running HF model with model args (hf model name in model_args as well):

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

Task formats

YAML+jinja, can run python code in some of the params.

task: coqa
dataset_path: EleutherAI/coqa
output_type: generate_until
training_split: train
validation_split: validation
doc_to_text: !function utils.doc_to_text
doc_to_target: !function utils.doc_to_target
process_results: !function utils.process_results
should_decontaminate: true
doc_to_decontamination_query: "{{story}} {{question.input_text|join('\n')}}"
generation_kwargs:
  until:
    - "\nQ:"
metric_list:
  - metric: em
    aggregation: mean
    higher_is_better: true
  - metric: f1
    aggregation: mean
    higher_is_better: true

Features

Supports HF models/datasets
Jinja templates, separate packages for IFeval, promptsource etc.
Many options for post-processing, answer extraction, fewshot etc. the usual, all live in the YAML

Inference

HF, incl. local, OpenAI API format, but other frameworks as well — nvidia nemo etc.
Llama.cpp / vLLM
multi-GPU with accelerate: accelerate launch -m lm_eval --model ...

Saving and caching

JSON through --output_path param
--log_samples logs samples
--use-cache caches stuff and reruns it only when needed
--hf_hub_log_args logs the results to HF ! (documentation broken though)

Misc

Last (most complex) version of HF Hub leaderboard (before it got discontinued): lm-evaluation-harness/lm_eval/tasks/leaderboard/README.md at main · EleutherAI/lm-evaluation-harness
Python integration: simple_evaluate(): lm-evaluation-harness/docs/interface.md at main · EleutherAI/lm-evaluation-harness

lmms-eval

EvolvingLMMs-Lab/lmms-eval: Accelerating the development of large multimodal models (LMMs) with one-click evaluation module - lmms-eval.
very close in spirit to eval-harness but multimodal

Sample¹:

export HF_HOME="~/.cache/huggingface"
export AZURE_OPENAI_API_KEY=""
export AZURE_OPENAI_API_BASE=""
export AZURE_OPENAI_API_VERSION="2023-07-01-preview"

# pip install git+https://github.com/EvolvingLMMs-Lab/lmms-eval.git
python3 -m lmms_eval \
    --model openai_compatible \
    --model_args model_version=gpt-4o-2024-11-20,azure_openai=True \
    --tasks mme,mmmu_val \
    --batch_size 1

Task yamls look very similar: lmms-eval/lmms_eval/tasks/gqa/gqa.yaml at main · EvolvingLMMs-Lab/lmms-eval

Evalverse

Evalverse | Evalverse Docs
Looks abandoned
One package to run multiple eval frameworks:
- lm-eval for usual benchmarks
- Chat: evaluated using FastChat llm-as-a-judge
  - MT-Bench
  - IFEval
  - EQ-Bench
Their params:Evaluation | Evalverse Docs
- also “Common Arguments” (hah)
- And then arguments for each benchmark separately

Architecture

Evaluator runs the library given by the Connector

Running

Pasted image 20250605144055.png (h6_en is lm-eval)

Results

JSON, one for each task.
They have a Reporter class that summarizes and analyzes the results saved by the evaluator

OpenAI evals

Link: openai/evals: Evals is a framework for evaluating LLMs and LLM systems, and an open-source registry of benchmarks.
I like their documentation!

Running

oaieval gpt-3.5-turbo test-match

Task formats

Data

Data has to be in JSONL format.

{
  "input": [
    {
      "role": "system",
      "content": "You are an assistant with knowledge of U.S. state laws. Answer the questions accurately."
    },
    {
      "role": "user",
      "content": "List the states where adultery is technically illegal. Only provide a list of states with no explanation."
    }
  ],
  "ideal": "Alabama, Arizona, Florida, Idaho, Illinois, Kansas, Michigan, Minnesota, Mississippi, New York, North Carolina, Oklahoma, Rhode Island, South Carolina, Virginia, Wisconsin, Georgia"
}

Building an eval

Registering the eval²:

<eval_name>:
  id: <eval_name>.dev.v0
  description: <description>
  metrics: [accuracy]

<eval_name>.dev.v0:
  class: evals.elsuite.basic.match:Match
  args:
    samples_jsonl: <eval_name>/samples.jsonl

Eval templates

evals/docs/eval-templates.md at main · openai/evals templates to make task creation easier
- e.g. basic: match/fuzzy-match/..
- model-graded: use a model to parse the completion into an easily parsable shape, e.g. yes/no/maybe.

Model-graded templates

Sample yaml:³

humor_likert:
  prompt: |-
    Is the following funny?

    {completion}

    Answer using the scale of 1 to 5, where 5 is the funniest.
  choice_strings: "12345"
  choice_scores: from_strings
  input_outputs:
    input: completion

closedqa:
  prompt: |-
    You are assessing a submitted answer on a given task based on a criterion. Here is the data:
    [BEGIN DATA]
    ***
    [Task]: {input}
    ***
    [Submission]: {completion}
    ***
    [Criterion]: {criteria}
    ***
    [END DATA]
    Does the submission meet the criterion? First, write out in a step by step manner your reasoning about the criterion to be sure that your conclusion is correct. Avoid simply stating the correct answers at the outset. Then print only the single character "Y" or "N" (without quotes or punctuation) on its own line corresponding to the correct answer. At the end, repeat just the letter again by itself on a new line.

    Reasoning:
  eval_type: cot_classify
  choice_scores:
    "Y": 1.0
    "N": 0.0
  choice_strings: 'YN'
  input_outputs:
    input: "completion"

Includes
- closedqa (is the answer given in the prompt?)
- battle.yaml: pairwise choice basically
Notably, I don’t really see a way to do multiple choice using their basic eval templates, a la lm-eval doc_to_choice.

Inference

Completion functions: evals/docs/completion-fns.md at main · openai/evals
generalization of “completion”, e.g. also to allow the model to open a browser or whatever

langchain/llm/text-davinci-003:
  class: evals.completion_fns.langchain_llm:LangChainLLMCompletionFn
  args:
    llm: OpenAI
    llm_kwargs:
      model_name: text-davinci-003

langchain/llm/flan-t5-xl:
  class: evals.completion_fns.langchain_llm:LangChainLLMCompletionFn
  args:
    llm: HuggingFaceHub
    llm_kwargs:
      repo_id: google/flan-t5-xl

Model types

Not immediately easily exposed, definitely supports OpenAI and LangChain and HF, but it’s not intuitive.

HF Lighteval

Lighteval

Running

Pasted image 20250605151104.png

lighteval accelerate \
     "model_name=openai-community/gpt2" \
     "leaderboard|truthfulqa:mc|0|0"

The syntax: {suite}|{task}|{num_few_shot}|{0 for strict num_few_shots, or 1 to allow a truncation if context size is too small}

Task formats

Python files: Adding a Custom Task.
Example template: lighteval/community_tasks/_template.py
German RAG eval: lighteval/community_tasks/german_rag_evals.py at main · huggingface/lighteval
Currently implemented tasks: lighteval/src/lighteval/tasks at main · huggingface/lighteval
Example task: lighteval/examples/custom_tasks_templates/custom_yourbench_task_mcq.py at main · huggingface/lighteval

    return Doc(
        instruction=ZEROSHOT_QA_INSTRUCTION,
        task_name=task_name,
        query=ZEROSHOT_QA_USER_PROMPT.format(question=line["question"], options=options),
        choices=line["choices"],
        gold_index=gold_index,
    )

yourbench_mcq = LightevalTaskConfig(
    name="HF_TASK_NAME",  # noqa: F821
    suite=["custom"],
    prompt_function=yourbench_prompt,
    hf_repo="HF_DATASET_NAME",  # noqa: F821
    hf_subset="lighteval",
    hf_avail_splits=["train"],
    evaluation_splits=["train"],
    few_shots_split=None,
    few_shots_select=None,
    generation_size=8192,
    metric=[Metrics.yourbench_metrics],
    trust_dataset=True,
    version=0,
)

Features

Inference

Model types

Many, and model configs are yamls: lighteval/examples/model_configs at main · huggingface/lighteval. For example: lighteval/examples/model_configs/litellm_model.yaml at main · huggingface/lighteval

model_parameters:
  model_name: "openai/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
  provider: "openai"
  base_url: "https://router.huggingface.co/hf-inference/v1"
  generation_parameters:
    temperature: 0.5
    max_new_tokens: 256
    top_p: 0.9
    seed: 0
    repetition_penalty: 1.0
    frequency_penalty: 0.0

Misc

Their default prompts: lighteval/src/lighteval/tasks/default_prompts.py at main · huggingface/lighteval

HELM

stanford-crfm/helm: Holistic Evaluation of Language Models (HELM)
Their famous leaderboard: Holistic Evaluation of Language Models (HELM)
Documentation: Index - CRFM HELM

Running

# Run benchmark 
helm-run --run-entries mmlu:subject=philosophy,model=openai/gpt2 --suite my-suite --max-eval-instances 10 
# Summarize benchmark results 
helm-summarize --suite my-suite

Task formats

Features

Inference

A model has metadata (description) and deployment (actually how to run it / implementation). Adding New Models - CRFM HELM Both are yamls.

HF model deployiment (but running locally!):

  - name: huggingface/gemma-2-9b-it
    model_name: google/gemma-2-9b-it
    tokenizer_name: google/gemma-2-9b
    max_sequence_length: 8192
    client_spec:
      class_name: "helm.clients.huggingface_client.HuggingFaceClient"
      args:
        device_map: auto
        torch_dtype: torch.bfloat16

I’m not certain how that connects to their Hugging Face Model Hub Integration - CRFM HELM — tl;dr only AutoModelForCausalLM. To run: helm-run \ --run-entries boolq:model=stanford-crfm/BioMedLM \ --enable-huggingface-models stanford-crfm/BioMedLM \ --suite v1 \ --max-eval-instances 10

All (many!) deployments: helm/src/helm/config/model_deployments.yaml at main · stanford-crfm/helm

vLLM example uses a OpenAI-compatible inference server

Pasted image 20250605152132.png

OpenCompass

Never heard of it but looks cool! And supports many types of evals.

Subjective Evaluation Guidance — OpenCompass 0.4.2 documentation

All configs in Python! models tasks etc.

Model types

All configs in python!

# model_cfg.py
from opencompass.models import HuggingFaceCausalLM

models = [
    dict(
        type=HuggingFaceCausalLM,
        path='huggyllama/llama-7b',
        model_kwargs=dict(device_map='auto'),
        tokenizer_path='huggyllama/llama-7b',
        tokenizer_kwargs=dict(padding_side='left', truncation_side='left'),
        max_seq_len=2048,
        max_out_len=50,
        run_cfg=dict(num_gpus=8, num_procs=1),
    )
]

OpenAI:

from opencompass.models import OpenAI

models = [
    dict(
        type=OpenAI,                             # Using the OpenAI model
        # Parameters for `OpenAI` initialization
        path='gpt-4',                            # Specify the model type
        key='YOUR_OPENAI_KEY',                   # OpenAI API Key
        max_seq_len=2048,                        # The max input number of tokens
        # Common parameters shared by various models, not specific to `OpenAI` initialization.
        abbr='GPT-4',                            # Model abbreviation used for result display.
        max_out_len=512,                         # Maximum number of generated tokens.
        batch_size=1,                            # The size of a batch during inference.
        run_cfg=dict(num_gpus=0),                # Resource requirements (no GPU needed)
    ),
]

DeepEval

confident-ai/deepeval: The LLM Evaluation Framework
basically unit-test for LLMs — optionally literally using pytest!
“large amount of metrics” usable with “ANY LLM of your choice”
- includes RAG tool use etc.
have their free cloud thingy to share reports etc., optional but they push it heavily. Something like wandb — visualize scores compare etc.
Supports evaluating LlamaIndex | DeepEval

Tasks

From their README:

import pytest
from deepeval import assert_test
from deepeval.metrics import GEval
from deepeval.test_case import LLMTestCase, LLMTestCaseParams

def test_case():
    correctness_metric = GEval(
        name="Correctness",
        criteria="Determine if the 'actual output' is correct based on the 'expected output'.",
        evaluation_params=[LLMTestCaseParams.ACTUAL_OUTPUT, LLMTestCaseParams.EXPECTED_OUTPUT],
        threshold=0.5
    )
    test_case = LLMTestCase(
        input="What if these shoes don't fit?",
        # Replace this with the actual output from your LLM application
        actual_output="You have 30 days to get a full refund at no extra cost.",
        expected_output="We offer a 30-day full refund at no extra costs.",
        retrieval_context=["All customers are eligible for a 30 day full refund at no extra costs."]
    )
    assert_test(test_case, [correctness_metric])

Models

the usual ones, including neat HF support k The vLLM bit describes using an openAI-compatible endpoints: vLLM | DeepEval - The Open-Source LLM Evaluation Framework, same as for LMStudio

deepeval set-local-model --model-name=<model_name> \  
--base-url="http://localhost:8000/v1/" \  
--api-key=<api-key>

Misc

Can do component-based evals through @observe decorators, “avoiding rewriting your app just for testing”
Supports many vector databases eval for RAG retrieval eval

Other stuff

FastChat MTBench

FastChat/fastchat/llm_judge at main · lm-sys/FastChat
pure Python, llm-as-a-judge
one python script to generate answers, another to get LLM judgements, a third to calculate results and draw plots — neat pattern
FastChat powers LLM Arena, and also provides an OpenAI-compatible API for the models it supports

Autoevals

braintrustdata/autoevals: AutoEvals is a tool for quickly and easily evaluating AI model outputs using best practices.
python+TS, no runner but framework
API-first, doesn’t try to run local inference, which again I like
Generally I really like it

Pydantic evals

Evals - PydanticAI
Yet another unit tests for LLM approach, flexible, supports LLM-as-a-judge

Nel mezzo del deserto posso dire tutto quello che voglio.

serhii.net

Evaluation harnesses list and notes

EleutherAI lm-eval-harness

Running

Task formats

Features

Inference

Saving and caching

Misc

lmms-eval

Evalverse

Architecture

Running

Results

OpenAI evals

Running

Task formats

Data

Building an eval

Eval templates

Model-graded templates

Inference

Model types

HF Lighteval

Running

Task formats

Features

Inference

Model types

Misc

HELM

Running

Task formats

Features

Inference

OpenCompass

Model types

DeepEval

Tasks

Models

Misc

Other stuff

FastChat MTBench

Autoevals

Pydantic evals