We have a revamp of the Evaluation Harness library internals staged on the [big-refactor](https://github.com/EleutherAI/lm-evaluation-harness/tree/big-refactor) branch! It is far along in progress, but before we start to move the `master` branch of the repository over to this new design with a new version release, we'd like to ensure that it's been tested by outside users and there are no glaring bugs.
We’d like your help to test it out! you can help by:
1. Trying out your current workloads on the big-refactor branch, and seeing if anything breaks or is counterintuitive,
2. Porting tasks supported in the previous version of the harness to the new YAML configuration format. Please check out our [task implementation guide](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/docs/new_task_guide.md) for more information.
If you choose to port a task not yet completed according to [our checklist](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/lm_eval/tasks/README.md), then you can contribute it by opening a PR containing [Refactor] in the name with:
- A command of the form `python -m lm_eval --model hf --model_args ..... --tasks <task name> ...` which will run the task in the `master` branch, and what the score is
- A command of the form `python -m lm_eval --model hf --model_args ..... --tasks <task name> ...` to run the task in your PR branch to `big-refactor`, and what the resulting score is, to show that we achieve equality between the two implementations.
Lastly, we'll no longer be accepting new feature requests beyond those that are already open to the master branch as we carry out this switch to the new version over the next week, though we will be accepting bugfixes to `master` branch and PRs to `big-refactor`. Feel free to reach out in the #lm-thunderdome channel of the EAI discord for more information.
## Overview
This project provides a unified framework to test generative language models on a large number of different evaluation tasks.
Features:
-Many tasks implemented, 200+ tasks [implemented in the old framework](https://github.com/EleutherAI/lm-evaluation-harness/blob/master/docs/task_table.md) which require porting to the new setup as described in [the new task guide](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/docs/new_task_guide.md).
-Over 60 standard academic benchmarks for LLMs, with hundreds of subtasks and variants implemented.
- Support for models loaded via [transformers](https://github.com/huggingface/transformers/)(including quantization via [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ)), [GPT-NeoX](https://github.com/EleutherAI/gpt-neox), and [Megatron-DeepSpeed](https://github.com/microsoft/Megatron-DeepSpeed/), with a flexible tokenization-agnostic interface.
- Support for commercial APIs including [OpenAI](https://openai.com), [goose.ai](https://goose.ai), and [TextSynth](https://textsynth.com/).
- Support for evaluation on adapters (e.g. LoRA) supported in [HuggingFace's PEFT library](https://github.com/huggingface/peft).
...
...
@@ -32,7 +18,7 @@ The Language Model Evaluation Harness is the backend for 🤗 Hugging Face's pop
## Install
To install the `lm-eval`refactor branch from the github repository, run:
To install the `lm-eval`package from the github repository, run:
@@ -58,7 +44,6 @@ To install the package with all extras, run
pip install-e".[all]"
```
## Support
The best way to get support is to open an issue on this repo or join the EleutherAI discord server](discord.gg/eleutherai). The `#lm-thunderdome` channel is dedicated to developing this project and the `#release-discussion` channel is for receiving support for our releases.
...
...
@@ -156,7 +141,7 @@ A full accounting of the supported and planned libraries + APIs can be seen belo
| API or Inference Server | Implemented? | `--model <xxx>` name | Models supported: | Request Types: |
We support wildcards in task names, for example you can run all of the machine-translated lambada tasks via `--task lambada_openai_mt_*`.
## Implementing new tasks
To implement a new task in the eval harness, see [this guide](./docs/new_task_guide.md).
As a start, we currently only support one prompt per task, which we strive to make the "standard" as defined by the benchmark's authors. If you would like to study how varying prompts causes changes in the evaluation score, we support prompts authored in the [Promptsource Library](https://github.com/bigscience-workshop/promptsource/tree/main) as described further in [the task guide](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/lm_eval/docs/new_task_guide.md) and [the advanced task guide](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/lm_eval/docs/advanced_task_guide.md) and welcome contributions of novel task templates and task variants.
## How to Contribute or Learn More?
...
...
@@ -245,35 +224,19 @@ For more information on the library and how everything fits together, check out
You can also ask for help, or discuss new features with the maintainers in the #lm-thunderdome channel of the EleutherAI discord! If you've used the library and have had a positive (or negative) experience, we'd love to hear from you!
### Implementing new tasks
To implement a new task in the eval harness, see [this guide](./docs/new_task_guide.md).
As a start, we currently only support one prompt per task, which we strive to make the "standard" as defined by the benchmark's authors. If you would like to study how varying prompts causes changes in the evaluation score, we support prompts authored in the [Promptsource Library](https://github.com/bigscience-workshop/promptsource/tree/main) as described further in [the task guide](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/lm_eval/docs/new_task_guide.md) and [the advanced task guide](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/lm_eval/docs/advanced_task_guide.md) and welcome contributions of novel task templates and task variants.
## Cite as
```
@software{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
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},
@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},
Where `Instance` is a dataclass defined in [`lm_eval.api.instance`](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/lm_eval/api/instance.py) with property `args`which returns a tuple of (context, continuation).
Where `Instance` is a dataclass defined in [`lm_eval.api.instance`](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/lm_eval/api/instance.py) with property `args`of request-dependent type signature described below.
We support three types of requests, consisting of different interactions / measurements with an autoregressive LM.
...
...
@@ -56,14 +56,37 @@ All three request types take as input `requests` of type `list[Instance]` that h
-`generate_until`
- Each request contains `Instance.args : Tuple[str, dict]` containing 1. an input string to the LM and 2. a dictionary of keyword arguments used to control generation parameters.
-
- Using this input and these generation parameters, text will be sampled from the language model (typically until a maximum output length or specific stopping string sequences--for example, `{"until": ["\n\n", "."], "max_gen_toks": 128}`).
- The generated input+output text from the model will then be returned.
-`loglikelihood`
-
- Each request contains `Instance.args : Tuple[str, str]` containing 1. an input string to the LM and 2. a target string on which the loglikelihood of the LM producing this target, conditioned on the input, will be returned.
- Each request will have, as result, `(ll, is_greedy): Tuple[float, int]` returned, where `ll` is a floating point number representing the log probability of generating the target string conditioned on the input, and `is_greedy` being either the value `0` or `1`, with it being `1` if and only if the target string *would be generated by greedy sampling from the LM* (that is, if the target string is the *most likely* N-token string to be output by the LM given the input. )
-`loglikelihood_rolling`, and args passed to it
-`loglikelihood_rolling`
- Each request contains `Instance.args : Tuple[str]`, which is an input string to the model whose *entire* loglikelihood, conditioned on purely the EOT token, will be calculated.
- This is used to evaluate *perplexity* on a data distribution.
- It should return `(ll,) : Tuple[float]` , a.k.a. solely the *loglikelihood* of producing each piece of text given no starting input.
To allow a model to be evaluated on all types of tasks, you will need to implement these three types of measurements (note that `loglikelihood_rolling` is a special case of `loglikelihood`). For a reference implementation, check out `lm_eval/models/huggingface.py` !
**Tip: be careful of indexing in loglikelihood!**
LMs take in tokens in position `[0 1 2 ... N]` and output a probability distribution for token position `N+1`. We provide a simplified graphic here, excerpted from `huggingface.py`:
```
# how this all works (illustrated on a causal decoder-only setup):
# CTX CONT
# inp 0 1 2 3|4 5 6 7 8 9 <- last token is deleted by inp[:, :-1]
# model \ \
# logits 1 2 3|4 5 6 7 8 9 <- the ctx half gets tossed out by the
The final token of the target is not passed into the LM, because we want the LM's predictions *up to but not past* that final target token. For more information, check out https://github.com/EleutherAI/lm-evaluation-harness/issues/942 .
## Registration
Congrats on implementing your model! Now it's time to test it out.
...
...
@@ -81,7 +104,9 @@ class MyCustomLM(LM):
Using this decorator results in the class being added to an accounting of the usable LM types maintained internally to the library at `lm_eval.api.registry.MODEL_REGISTRY`. See `lm_eval.api.registry` for more detail on what sorts of registries and decorators exist in the library!
## Testing
We also recommend that new model contributions be accompanied by short tests of their 3 core functionalities, at minimum. To see an example of such tests, look at https://github.com/EleutherAI/lm-evaluation-harness/blob/35bdecd379c0cefad6897e67db892f4a6026a128/tests/test_ggml.py .
As a concrete example, we'll walk through reimplementing the `gsm8k` benchmark (a *generative* task which requires sampling text from a model) and the `sciq` benchmark. (a *discriminative*, or *multiple choice*, task where the model picks the most likely of several fixed answer choices).
In this document, we'll walk through the basics of implementing a static benchmark evaluation in two formats: a *generative* task which requires sampling text from a model, such as [`gsm8k`](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/lm_eval/tasks/gsm8k/gsm8k.yaml), and a *discriminative*, or *multiple choice*, task where the model picks the most likely of several fixed answer choices, such as [`sciq`](https://github.com/EleutherAI/lm-evaluation-harness/blob/big-refactor/lm_eval/tasks/sciq/sciq.yaml).
## Creating a YAML file
...
...
@@ -45,6 +45,16 @@ dataset_name: ... # the dataset configuration to use. Leave `null` if your datas
dataset_kwargs:null# any extra keyword arguments that should be passed to the dataset constructor, e.g. `data_dir`.
```
------------------------------
**Tip:** To load a local dataset for evaluation, you can specify data files in the `dataset_kwargs` field, such as the following for JSON files:
```
dataset_path: json
dataset_name: null
dataset_kwargs:
data_files: /path/to/my/json
```
-------------------------------
Next, we'd like to tell our task what the dataset's train, validation, and test splits are named, if they exist:
```yaml
...
...
@@ -116,7 +126,7 @@ doc_to_choice: ['No', 'Yes']
We support the [Jinja 2](https://jinja.palletsprojects.com/en/3.1.x/) templating language for writing prompts. In practice, this means you can take your dataset's columns and do many basic string manipulations to place each document into prompted format.
Take for example `super_glue/boolq`, as input, we'd like to use the features `passage` and `question` and string them together so that for a a sample line `doc`, the model sees something the format of:
Take for example the dataset `super_glue/boolq`. As input, we'd like to use the features `passage` and `question` and string them together so that for a a sample line `doc`, the model sees something the format of:
```
doc["passage"]
Question: doc["question"]?
...
...
@@ -285,7 +295,7 @@ It's now time to check models' performance on your task! In the evaluation harne
To enable this, we provide a checklist that should be completed when contributing a new task, to enable accurate book-keeping and to ensure that tasks added to the library are well-tested and, where applicable, precedented.
f'Both target_delimiter and target choice: "{choice}" does not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" do not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
