Home

Awesome

lm-evaluation-harness + promptsource

codecov

Overview

This project provides a unified framework to test causal (GPT-2, GPT-3, GPTNeo, etc) and seq2seq (T5, T0) language models via prompt evaluation.

As of now, all prompts are provided via the promptsource eval-hackathon branch; all datasets are from huggingface datasets.

This fork is not backwards compatible with the original evaluation harness.

Installation

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

CLI Usage 🖥️

To evaluate a model (e.g. GPT-2) on NLP tasks such as SuperGLUE WiC, you can run the following command:

python main.py \
    --model_api_name 'hf-causal' \
    --model_args pretrained='gpt2' \
    --task_name 'wic' \
    --template_names 'same_sense','polysemous' \
    --device cpu

Additional arguments can be provided to the model constructor using the --model_args flag. For larger models supported by HuggingFace transformers, we provide parallelism and mixed-precision utilities through the accelerate package. It can be activated for hf-causal/hf-seq2seq by passing use_accelerate=True and dtype=half to the --model_args flag, respectively. For finer grained control over accelerate options, see the constructor docstrings for HuggingFaceAutoLM in huggingface.py.

python main.py \
    --model_api_name 'hf-causal' \
    --model_args use_accelerate=True,pretrained='facebook/opt-13b' \
    --task_name wnli

If you have access to the OpenAI API, you can also evaluate GPT-3 engines:

export OPENAI_API_SECRET_KEY={YOUR_KEY_HERE}
python main.py \
    --model_api_name 'openai' \
    --model_args engine='curie' \
    --task_name hans

When reporting results from eval harness, please include the task versions (shown in results["versions"]) for reproducibility. This allows bug fixes to tasks while also ensuring that previously reported scores are reproducible.

Detailed Usage

usage: main.py [-h] --model_api_name MODEL_API_NAME [--model_args MODEL_ARGS] --task_name TASK_NAME
               [--template_names TEMPLATE_NAMES] [--num_fewshot NUM_FEWSHOT] [--batch_size BATCH_SIZE]
               [--device DEVICE] [--limit LIMIT] [--output_path OUTPUT_PATH] [--template_idx TEMPLATE_IDX]
               [--bootstrap_iters BOOTSTRAP_ITERS] [--no_tracking] [--use_cache]

optional arguments:
  -h, --help            show this help message and exit
  --model_api_name MODEL_API_NAME
                        Name of the model API to use. See `lm_eval.list_model_apis()` for available APIs
  --model_args MODEL_ARGS
                        Model constructor args that you'd pass into a model of type `--model_api_name`. These must
                        be comma-separated keyword args, e.g. `key1=value1,key2=value2`, with no spaces
  --task_name TASK_NAME
                        Name of the task to use as found in the lm_eval registry. See: `lm_eval.list_tasks()`
  --task_args TASK_ARGS
                        Optional task constructor args that you'd pass into a task class of kind " `--task_name`.
                        These must be comma-separated keyword args, e.g. `key1=value1,key2=value2`, with no spaces.
                        WARNING: To avoid parsing errors, ensure your strings are quoted. For example,
                            `example_separator='\n+++\n'`
                        WARNING: Values must NOT contain commas.
  --template_names TEMPLATE_NAMES
                        Comma-separated list of template names for the specified task. Example:
                        `> python main.py ... --task_name rte --template_names imply,mean`
                        - Default: `all_templates`
                        - General Selectors:
                            - `"all_templates"`: Selects all templates for the task
                            - `"original_templates"`: Selects only templates that are designed to match the original task
  --num_fewshot NUM_FEWSHOT
  --batch_size BATCH_SIZE
  --seed SEED
  --device DEVICE       The device to place your model onto, e.g. cuda:0. For large models available through the
                        HuggingFace Hub you should use `accelerate` by passing `use_accelerate=True` to
                        `--model_args`
  --limit LIMIT         Limit the number of examples to evaluate on; ONLY USE THIS FOR DEBUGGING PURPOSES
  --output_path OUTPUT_PATH
                        Use output_path as `output_filename`. For example:
                        `> python main.py ... --output_path blop`
                        # saves files into `outputs/blop.json` Warning: You currently cannot change/add folder
                        structure.
  --template_idx TEMPLATE_IDX
                        Choose template by index from available templates
  --bootstrap_iters BOOTSTRAP_ITERS
                        Iters for stderr computation
  --no_tracking         Skip carbon emission tracking
  --use_cache           Whether to cache your model's predictions or not

Library Usage 📖

You can also use lm_eval as a library:

import lm_eval

model = lm_eval.get_model("hf-causal", pretrained="gpt2", device="cpu")
tasks = lm_eval.get_task_list(
    "superglue_rte",
    template_names=["does this imply", "must be true"])
results = lm_eval.evaluate(model=model, tasks=tasks)

The main user-facing functions are:

Some high-level convenience functions are also made available:

Gotchas 🩹

Features

Implementing new tasks

To implement a new task in eval harness, follow the PromptSourceTask template.

Using load_from_disk instead of load_dataset

You can use load_from_disk (convenient on Jean Zay supercomputer) by setting task_args download_mode='load_from_disk',data_dir=<data/path>