LLamaRecipes/tools/benchmarks/llm_eval_harness/meta_eval_reproduce/README.md

Reproducing Meta 3.1 Evaluation Metrics Using LM-Evaluation-Harness

As Meta Llama models gain popularity, evaluating these models has become increasingly important. We have released all the evaluation details for Meta-Llama 3.1 models as datasets in the 3.1 evals Hugging Face collection. This recipe demonstrates how to closely reproduce the Llama 3.1 reported benchmark numbers using the lm-evaluation-harness library and our prompts from the 3.1 evals datasets on selected tasks.

Disclaimer

1. This recipe is not the official implementation of Meta Llama evaluation. It is based on public third-party libraries, as this implementation is not mirroring Meta Llama evaluation, this may lead to minor differences in the reproduced numbers.
2. Model Compatibility: This tutorial is specifically for Llama 3 based models, as our prompts include Meta Llama 3 special tokens, e.g. <|start_header_id|>user<|end_header_id|>. It will not work with models that are not based on Llama 3.

Insights from Our Evaluation Process

There are 4 major differences in terms of the eval configurations and prompting methods between this implementation and Hugging Face leaderboard implementation.

• Prompts: We use Chain-of-Thought(COT) prompts while Hugging Face leaderboard does not. The prompts that define the output format are also different.
• Metric calculation: For MMLU-Pro, BBH, GPQA tasks, we ask the model to generate response and score the parsed answer from generated response, while Hugging Face leaderboard evaluation is comparing log likelihood of all label words, such as [ (A),(B),(C),(D) ].
• Parsers: For generative tasks, where the final answer needs to be parsed before scoring, the parser functions can be different between ours and Hugging Face leaderboard evaluation, as our prompts that define the model output format are designed differently.
• Inference: We use internal LLM inference solution that loads pytorch checkpoints and do not use padding, while Hugging Face leaderboard uses Hugging Face format model and sometimes will use padding depending on the tasks type and batch size.
• Tasks We run benchmarks on BBH and MMLU-Pro only for pretrained models and Math-Hard, IFeval, GPQA, only for pretrained models.

Given those differences, our reproduced number can not be compared to the numbers in the Hugging Face Open LLM Leaderboard v2, even if the task names are the same.

Environment setups

Please install lm-evaluation-harness and our llama-recipe repo by following:

pip install lm-eval[math,ifeval,sentencepiece,vllm]==0.4.3
git clone git@github.com:meta-llama/llama-recipes.git
cd llama-recipes
pip install -U pip setuptools
pip install -e .
cd tools/benchmarks/llm_eval_harness/meta_eval_reproduce

To access our 3.1 evals Hugging Face collection, you must:

• Log in to the Hugging Face website and click the 3.1 evals dataset pages and agree to the terms.
• Follow the Hugging Face authentication instructions to gain read access for your machine.

It is recommended to read the dataset card to understand the meaning of each column and use the viewer feature in the Hugging Face dataset to view our dataset. It is important to have some basic understanding of our dataset format and content before proceeding.

Task Selection

Given the extensive number of tasks available (12 for pretrained models and 30 for instruct models), here we will focus on tasks that overlap with the popular Hugging Face Open LLM Leaderboard v2 as shown in the following:

• Tasks for pretrained models: BBH and MMLU-Pro
• Tasks for instruct models: Math-Hard, IFeval, GPQA, and MMLU-Pro

Here, we aim to reproduce the Meta reported benchmark numbers on the aforementioned tasks using Hugging Face leaderboard implementation. Please follow the instructions below to make necessary modifications to use our eval prompts and reproduce our reported metrics.

Run eval tasks

1. We created eval_config.yaml to store all the arguments and hyperparameters. This is the main config file you need to change if you want to eval other models, and a part of eval_config.yaml looks like this:

model_name: "meta-llama/Meta-Llama-3.1-8B-Instruct" # The name of the model to evaluate. This must be a valid Meta Llama 3 based model name in the HuggingFace model hub."

evals_dataset: "meta-llama/Meta-Llama-3.1-8B-Instruct-evals" # The name of the 3.1 evals dataset to evaluate, please make sure this eval dataset corresponds to the model loaded. This must be a valid Meta Llama 3.1 evals dataset name in the Llama 3.1 Evals collection.
# Must be one of the following ["meta-llama/Meta-Llama-3.1-8B-Instruct-evals","meta-llama/Meta-Llama-3.1-70B-Instruct-evals","meta-llama/Meta-Llama-3.1-405B-Instruct-evals","meta-llama/Meta-Llama-3.1-8B-evals","meta-llama/Meta-Llama-3.1-70B-evals","meta-llama/Meta-Llama-3.1-405B-evals"]

tasks: "meta_instruct" # Available tasks for instruct model: "meta_math_hard", "meta_gpqa", "meta_mmlu_pro_instruct", "meta_ifeval"; or just use "meta_instruct" to run all of them.
# Available tasks for pretrain model: "meta_bbh", "meta_mmlu_pro_pretrain"; or just use "meta_pretrain" to run all of them.

tensor_parallel_size: 1 # The VLLM argument that speicify the tensor parallel size for the model, eg how many GPUs to use for a model copy.

data_parallel_size: 4 # The VLLM argument that speicify the data parallel size for the model, eg how copies of model will be used.

...

Change model_name to the model name you want to eval on and change the evals_dataset according to the model type and parameters. Remember to adjust the tensor_parallel_size to 2 or more to load the 70B models and change the data_parallel_size accordingly so that tensor_parallel_size * data_parallel_size is the number of GPUs you have. Please read the comments inside this yaml for detailed explanations on other parameters.

1. We already included all the related eval task yaml and python files in the meta_template folder, which define all the task implementation. You do not need to change those manually, we will use prepare_meta_eval.py to automatically change them later.

2. Then we can run prepare_meta_eval.py that reads the configuration from eval_config.yaml, copies everything in the template folder to a working folder work_dir, makes modification to those templates accordingly, prepares dataset if needed and prints out the CLI command to run the lm_eval.

To run the prepare_meta_eval.py, we can do:

python prepare_meta_eval.py --config_path ./eval_config.yaml

This script will load the default eval_config.yaml config and print out a lm_eval command to run meta_instruct group tasks, which includes meta_ifeval, meta_math_hard, meta_gpqa and meta_mmlu_pro_instruct, for meta-llama/Meta-Llama-3.1-8B-Instruct model using meta-llama/Meta-Llama-3.1-8B-Instruct-evals dataset.

An example output from prepare_meta_eval.py looks like this:

lm_eval --model vllm --model_args pretrained=meta-llama/Meta-Llama-3.1-8B-Instruct,tensor_parallel_size=1,dtype=auto,gpu_memory_utilization=0.9,data_parallel_size=4,max_model_len=8192,add_bos_token=True,seed=42 --tasks meta_instruct --batch_size auto --output_path eval_results --include_path ./work_dir --seed 42  --log_samples

1. Then just copy the lm_eval command printed by prepare_meta_eval.py back to your terminal and run it to get our reproduced result, which will be saved into eval_results folder by default.

NOTE: As for --model vllm, here we will use VLLM inference instead of Hugging Face inference because of the padding issue. By default, for the generative tasks, the lm-eval --model_args="{...}" --batch_size=auto command will use Hugging Face inference solution that uses a static batch method with left padding using EOS_token for Llama models, but our internal evaluation will load python original checkpoints and handle individual generation request asynchronously without any padding. To simulate this, we will use VLLM inference solution to do dynamic batching without any padding.

NOTE: As for add_bos_token=True, since our prompts in the evals dataset has already included all the special tokens required by instruct model, such as <|start_header_id|>user<|end_header_id|>, we will not use --apply_chat_template argument for instruct models anymore. However, we need to use add_bos_token=True flag to add the BOS_token back during VLLM inference, as the BOS_token is removed by default in this PR.

NOTE: For meta_math_hard tasks, some of our internal math ground truth has been converted to scientific notation, e.g. 6\sqrt{7} has been converted to 1.59e+1, which will be later handled by our internal math evaluation functions. As the lm-evaluation-harness math evaluation utils.py can not fully handle those conversion, we will use the original ground truth from the original dataset lighteval/MATH-Hard by joining the tables on the original input questions. The get_math_data function in the prepare_meta_eval.py will handle this step and produce a local parquet dataset file.

Moreover, we have modified this math_hard/utils.py to address two issues:

1. This python script only use a regular expression "Final Answer: The final answer is(.*?). I hope it is correct." to get the final answer, because this format is shown in the previous 4 shot examples prompts. However, our MATH Hard task is using 0 shot COT prompts that ask model to put the final answer into this string format Therefore, the final answer is: $\\boxed{answer}$. I hope it is correct. which can not be captured by previous regular expression, so we will use \\box{} to parse the final answer instead.

2. The is_equiv(x1: str, x2: str) function failed parse some ground truth, as we noticed some error logs like [utils.py:158] couldn't parse one of [0,1) or [0,1), so all those questions will be marked as wrong. We raised a issue to lm_evaluation_harness about this problem and will add a string equality check statement before going to is_equiv() function for now as a temporary solution.

NOTE: For meta_ifeval tasks, we have to use the original configs, such as instruction_id_list, kwargs, from wis-k/instruction-following-eval in order to use lm-evaluation-harness IFeval evaluation. We will perform similar join back method using get_ifeval_data function in the prepare_meta_eval.py to get a local parquet dataset file.

Yaml Config Deep Dive

Here, we will use MMLU-Pro as a example to show the steps to create a yaml config with detailed explanations, so that people can follow this example to create other tasks configurations if they want. For more information, please read lm-evaluation-harness new task guide.

1.Define the config to load datasets

We can use our 3.1 evals dataset as the source dataset and the corresponding subset and define the test split to latest. For example, if we want to reproduce the MMLU_Pro metric for 3.1 8B instruct, the following configs are needed as explained below:

task: meta_mmlu_pro_instruct
dataset_path: meta-llama/Meta-Llama-3.1-8B-Instruct-evals
dataset_name: Meta-Llama-3.1-8B-Instruct-evals__mmlu_pro__details
test_split: latest

If you want to run evaluation on 70B-Instruct, then it is recommended to change the dataset_path and dataset_name from 8B to 70B, even though 70B-instruct and 8B-instruct share the same prompts, the is_correct column, which can be used to get the difference between current reproduced result and the reported results for each sample, is different.

Note: Config files for Meta-Llama-3.1-8B-Instruct are already provided in each task subfolder under meta_template folder. Remember to change the eval dataset name according to the model type and DO NOT use pretrained evals dataset on instruct models or vice versa.

2.Configure preprocessing, prompts and ground truth

Here is the example yaml snippet in the MMLU-Pro that handles dataset preprocess, prompts and ground truth.

process_docs: !function utils.process_docs
doc_to_text: !function utils.doc_to_text
doc_to_target: gold

• process_docs : Defines the preprocess function for our datasets. In this case, we uses the process_docs python function that is defined in utils.py. This function will take the original dataset and output a processed dataset that has a out_doc, which contains problem which is the input question, gold which is the ground truth. We also renamed the is_correct column to previously_is_correct to allow detailed comparison for the difference of each sample between previously reported score and the reproduced score. You must use eval dataset and model with same parameters and same model type to get a valid comparison.

• doc_to_text: Defines the prompts. In the MMLU-Pro case, the input_final_prompts column always contains a list of a prompt, so we just use a python function that returns input_final_prompts[0].

• doc_to_target Defines the ground truth, which in the MMLU-Pro case, is the gold that derived from input_correct_responses[0].

3.Configure task type and parser

While Open LLM Leaderboard v2 uses multiple choice format for MMLU-Pro, BBH, GPQA tasks by comparing log likelihood of all label words, such as [ (A),(B),(C),(D) ], we use generative task option, by asking the model to generate response in sentences given our carefully designed prompts, then using some parsers to grab the final answer, and scoring that final answer based on the ground truth. Here is a example config in the MMLU-Pro that enable the generative task and defines the regex parser:

filter_list:
  - name: "strict-match"
    filter:
      - function: "regex"
        group_select: -1
        regex_pattern: 'best answer is ([A-Z])'
      - function: "take_first"

Since the MMLU-Pro task uses a 5-shot Chain-of-Thought(COT) prompts and the prompts are designed with explicitly instruction: "Your response should end with \"The best answer is [the_answer_letter].\" where the [the_answer_letter] is a letter from the provided choices.", we will use a simple and intuitive regex expression best answer is ([A-Z]) to parse the model response and take the last appearance as the final answer and this final answer will be scored based on the ground truth gold using exact match method.

4.Define generation and metric config

Then we need to define the generation and metric config, which looks like this:

generation_kwargs:
  until: []
  do_sample: false
  temperature: 0
  max_gen_toks: 1024
num_fewshot: 0
metric_list:
  - metric: exact_match
    aggregation: mean
    higher_is_better: true
    ignore_case: true
    ignore_punctuation: true

Here we set the num_fewshot to 0 as our prompts have already been converted to 5-shots, and the model generation will only stop if the generated output tokens exceeds 1024, as stated in the mmlu-pro eval details. We will set the do_sample to false and temperature to 0 as stated in our eval_config column in the dataset. We will use metric exact_match for this tasks and calculate the mean as our task aggregated number.

NOTE: While we tried our best to create the template files, those configs and functions are created based on public third-party library and are not exactly the same as our internal implementation, so there is a chance that the reproduced numbers are slightly different.

Results

Here is the comparison between our reported numbers and the reproduced numbers in this tutorial:

Model	MATH_HARD	GPQA_RAW	MMLU_PRO_RAW	IFeval
3.1 8B-Instruct reported	0.254	0.328	0.47	0.804
3.1 8B-Instruct reproduced	0.2424	0.3259	0.4675	0.7782
3.1 70B-Instruct reported	0.438	0.467	0.651	0.875
3.1 70B-Instruct reproduced	0.4388	0.4799	0.6475	0.848

Model	BBH_RAW	MMLU_PRO_RAW
3.1 8B reported	0.642	0.356
3.1 8B reproduced	0.6515	0.3572
3.1 70B reported	0.816	0.52
3.1 70B reproduced	0.8191	0.5225

From the table above, we can see that most of our reproduced results are very close to our reported number in the Meta Llama website.

NOTE: We used the average of inst_level_strict_acc,none and prompt_level_strict_acc,none to get the final number for IFeval as stated here.

NOTE: In the Meta Llama website, we reported the macro_avg metric, which is the average of all subtask average score, for MMLU-Protask, but here we are reproducing the micro_avg metric, which is the average score for all the individual samples, and those micro_avg numbers can be found in the eval_details.md.

NOTE: The reproduced numbers may be slightly different, as we observed around ±0.01 differences between each reproduce run because the latest VLLM inference is not very deterministic even with temperature=0. This behavior maybe related this issue. or it is expected due to 16-bits inference as stated in this comment and this comment.

Acknowledgement

This tutorial is inspired by leaderboard tasks implementation on the lm_evaluation_harness created by Hugging Face Open LLM Leaderboard v2 team. We also extend our gratitude to the lm-evaluation-harness github repo from EleutherAI.