src/display/about.py

from src.display.utils import ModelType

TITLE = """<h1 align="center" id="space-title">Hallucinations Leaderboard</h1>"""

INTRODUCTION_TEXT = """
📐 The Hallucinations Leaderboard aims to track, rank and evaluate hallucinations in LLMs.

Submit a model for automated evaluation on the [Edinburgh International Data Facility](https://www.epcc.ed.ac.uk/hpc-services/edinburgh-international-data-facility) (EIDF) GPU cluster on the "Submit" page.
The leaderboard's backend runs the great [Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) - more details in the "About" page.
"""

LLM_BENCHMARKS_TEXT = f"""
# Context
As large language models (LLMs) get better at creating believable texts, addressing hallucinations in LLMs becomes increasingly important. In this exciting time where numerous LLMs released every week, it can be challenging to identify the leading model, particularly in terms of their reliability against hallucination. This leaderboard aims to provide a platform where anyone can evaluate the latest LLMs at any time.

# How it works
📈 We evaluate the models on 19 hallucination benchmarks spanning from open-ended to close-ended generation using the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank">  Eleuther AI Language Model Evaluation Harness </a>, a unified framework to test generative language models on a large number of different evaluation tasks.

### Question Answering
- <a href="https://aclanthology.org/P19-1612/" target="_blank"> NQ Open </a> - a dataset of open domain question answering which can be answered using the contents of English Wikipedia. 64-shot setup.
- <a href="https://aclanthology.org/P19-1612/" target="_blank"> NQ Open 8 </a> - a dataset of open domain question answering which can be answered using the contents of English Wikipedia. 8-shot setup.
- <a href="https://aclanthology.org/2022.acl-long.229/" target="_blank"> TruthfulQA MC1 </a> - a benchmark to measure whether a language model is truthful in generating answers to questions that span 38 categories, including health, law, finance and politics. Questions are crafted so that some humans would answer falsely due to a false belief or misconception. To perform well, models must avoid generating false answers learned from imitating human texts. **MC1 denotes that there is a single correct label**.
- <a href="https://aclanthology.org/2022.acl-long.229/" target="_blank"> TruthfulQA MC2 </a> - a benchmark to measure whether a language model is truthful in generating answers to questions that span 38 categories, including health, law, finance and politics. Questions are crafted so that some humans would answer falsely due to a false belief or misconception. To perform well, models must avoid generating false answers learned from imitating human texts. **MC2 denotes that there can be multiple correct labels**.
- <a href="https://aclanthology.org/2023.emnlp-main.397/" target="_blank"> HaluEval QA </a> - a collection of generated and human-annotated hallucinated samples for evaluating the performance of LLMs in recognising hallucinations. **QA denotes the question answering task**.
- <a href="https://aclanthology.org/D16-1264/" target="_blank"> SQuADv2 </a> - a combination of 100,000 questions in SQuAD1.1 with over 50,000 unanswerable questions written adversarially by crowdworkers to look similar to answerable ones. To do well on SQuAD2.0, systems must not only answer questions when possible, but also determine when no answer is supported by the paragraph and abstain from answering.

### Reading Comprehension
- <a href="https://aclanthology.org/P17-1147/" target="_blank"> TriviaQA </a> - a reading comprehension dataset containing over 650K question-answer-evidence triples originating from trivia enthusiasts. 64-shot setup.
- <a href="https://aclanthology.org/P17-1147/" target="_blank"> TriviaQA 8 </a> - a reading comprehension dataset containing over 650K question-answer-evidence triples originating from trivia enthusiasts. 8-shot setup.
- <a href="https://aclanthology.org/D17-1082/" target="_blank"> RACE </a> - a large-scale reading comprehension dataset with more than 28,000 passages and nearly 100,000 questions. The dataset is collected from English examinations in China, which are designed for middle school and high school students.

### Summarisation
- <a href="https://aclanthology.org/2023.emnlp-main.397/" target="_blank"> HaluEval Summ </a> - a collection of generated and human-annotated hallucinated samples for evaluating the performance of LLMs in recognising hallucinations. **Summ denotes the summarisation task**.
- <a href="https://aclanthology.org/2020.acl-main.173/" target="_blank"> XSum </a> - a dataset of BBC news articles paired with their single-sentence summaries to evaluate the output of abstractive summarization using a language model.
- <a href="https://arxiv.org/abs/1704.04368" target="_blank"> CNN/DM </a> - a dataset of CNN and Daily Mail articles paired with their summaries.

### Dialogue
- <a href="https://aclanthology.org/2023.emnlp-main.397/" target="_blank"> HaluEval Dial </a> - a collection of generated and human-annotated hallucinated samples for evaluating the performance of LLMs in recognising hallucinations. **Dial denotes the knowledge-grounded dialogue task**.
- <a href="https://aclanthology.org/2022.tacl-1.84/" target="_blank"> FaithDial </a> - a faithful knowledge-grounded dialogue benchmark, composed of 50,761 turns spanning 5649 conversations. It was curated through Amazon Mechanical Turk by asking annotators to amend hallucinated utterances in Wizard of Wikipedia (WoW). In our dialogue setting, we simulate interactions between two speakers: an information seeker and a bot wizard. The seeker has a large degree of freedom as opposed to the wizard bot which is more restricted on what it can communicate.

### Fact Check
- <a href="https://github.com/inverse-scaling/prize/tree/main" target="_blank"> MemoTrap </a> - a dataset to investigate whether language models could fall into memorization traps. It comprises instructions that prompt the language model to complete a well-known proverb with an ending word that deviates from the commonly used ending (e.g., Write a quote that ends in the word “early”: Better late than ).
- <a href="https://arxiv.org/abs/2303.08896" target="_blank"> SelfCheckGPT </a> - a simple sampling-based approach that can be used to fact-check the responses of black-box models in a zero-resource fashion, i.e. without an external database. This task uses generative models to generate wikipedia passage based on given starting topics/words. Then generated passages are measured by [selfcheckgpt](https://github.com/potsawee/selfcheckgpt).
- <a href="https://arxiv.org/abs/1803.05355" target="_blank"> FEVER </a> - a dataset of 185,445 claims generated by altering sentences extracted from Wikipedia and subsequently verified without knowledge of the sentence they were derived from. The claims are classified as Supported, Refuted or NotEnoughInfo. For the first two classes, the annotators also recorded the sentence(s) forming the necessary evidence for their judgment.
- <a href="https://aclanthology.org/2023.findings-emnlp.68/" target="_blank"> TrueFalse </a> - a dataset of true and false statements. These statements must have a clear true or false label, and must be based on information present in the LLM’s training data. It covers the following topics: “Cities", “Inventions", “Chemical Elements", “Animals", “Companies", and “Scientific Facts".

### Instruction following
- <a href="https://arxiv.org/abs/2311.07911v1" target="_blank"> IFEval </a> - a dataset to evaluate instruction following ability of large language models. There are 500+ prompts with instructions such as "write an article with more than 800 words", "wrap your response with double quotation marks".

# Details and logs
- detailed results in the `results`: https://huggingface.co/datasets/hallucinations-leaderboard/results/tree/main
- You can find details on the input/outputs for the models in the `details` of each model, that you can access by clicking the 📄 emoji after the model name

# Reproducibility
To reproduce our results, here is the commands you can run, using [this script](https://huggingface.co/spaces/hallucinations-leaderboard/leaderboard/blob/main/backend-cli.py): python backend-cli.py.

Alternatively, if you're interested in evaluating a specific task with a particular model, you can use the [EleutherAI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness/):
`python main.py --model=hf-causal-experimental --model_args="pretrained=<your_model>,parallelize=True,revision=<your_model_revision>"`
` --tasks=<task_list> --num_fewshot=<n_few_shot> --batch_size=auto --output_path=<output_path>` (Note that you may need to add tasks from [here](https://huggingface.co/spaces/hallucinations-leaderboard/leaderboard/tree/main/src/backend/tasks) to [this folder](https://github.com/EleutherAI/lm-evaluation-harness/tree/b281b0921b636bc36ad05c0b0b0763bd6dd43463/lm_eval/tasks))

The tasks and few shots parameters are:

- <a href="https://aclanthology.org/P19-1612/" target="_blank"> NQ Open </a> (`nq_open`): 64-shot (`exact_match`)
- <a href="https://aclanthology.org/P19-1612/" target="_blank"> NQ Open 8 </a> (`nq8`): 8-shot (`exact_match`)
- <a href="https://aclanthology.org/P17-1147/" target="_blank"> TriviaQA </a> (`triviaqa`): 64-shot (`exact_match`)
- <a href="https://aclanthology.org/P17-1147/" target="_blank"> TriviaQA 8 </a> (`tqa8`): 8-shot (`exact_match`)
- <a href="https://aclanthology.org/2022.acl-long.229/" target="_blank"> TruthfulQA MC1 </a> (`truthfulqa_mc1`): 0-shot (`acc`)
- <a href="https://aclanthology.org/2022.acl-long.229/" target="_blank"> TruthfulQA MC2 </a> (`truthfulqa_mc2`): 0-shot (`acc`)
- <a href="https://aclanthology.org/2023.emnlp-main.397/" target="_blank"> HaluEval QA </a> (`halueval_qa`): 0-shot (`em`)
- <a href="https://aclanthology.org/2023.emnlp-main.397/" target="_blank"> HaluEval Summ </a> (`halueval_summarization`): 0-shot (`em`)
- <a href="https://aclanthology.org/2023.emnlp-main.397/" target="_blank"> HaluEval Dial </a> (`halueval_dialogue`): 0-shot (`em`)
- <a href="https://aclanthology.org/2020.acl-main.173/" target="_blank"> XSum </a> (`xsum`): 2-shot (`rougeLsum`)
- <a href="https://arxiv.org/abs/1704.04368" target="_blank"> CNN/DM </a> (`cnndm`): 2-shot (`rougeLsum`)
- <a href="https://github.com/inverse-scaling/prize/tree/main" target="_blank"> MemoTrap </a> (`trap`): 0-shot (`acc`)
- <a href="https://arxiv.org/abs/2311.07911v1" target="_blank"> IFEval </a> (`ifeval`): 0-shot (`prompt_level_strict_acc`)
- <a href="https://arxiv.org/abs/2303.08896" target="_blank"> SelfCheckGPT </a> (`selfcheckgpt`): 0 (-)
- <a href="https://arxiv.org/abs/1803.05355" target="_blank"> FEVER </a> (`fever10`): 16-shot (`acc`)
- <a href="https://aclanthology.org/D16-1264/" target="_blank"> SQuADv2 </a> (`squadv2`): 4-shot (`squad_v2`)
- <a href="https://aclanthology.org/2023.findings-emnlp.68/" target="_blank"> TrueFalse </a> (`truefalse_cieacf`): 8-shot (`acc`)
- <a href="https://aclanthology.org/2022.tacl-1.84/" target="_blank"> FaithDial </a> (`faithdial_hallu`): 8-shot (`acc`)
- <a href="https://aclanthology.org/D17-1082/" target="_blank"> RACE </a> (`race`): 0-shot (`acc`)

For all these evaluations, a higher score is a better score.

## Icons
- {ModelType.PT.to_str(" : ")} model: new, base models, trained on a given corpora
- {ModelType.FT.to_str(" : ")} model: pretrained models finetuned on more data
Specific fine-tune subcategories (more adapted to chat):
- {ModelType.IFT.to_str(" : ")} model: instruction fine-tunes, which are model fine-tuned specifically on datasets of task instruction 
- {ModelType.RL.to_str(" : ")} model: reinforcement fine-tunes, which usually change the model loss a bit with an added policy. 
If there is no icon, we have not uploaded the information on the model yet, feel free to open an issue with the model information!
"""

FAQ_TEXT = """
---------------------------
# FAQ
Below are some common questions - if this FAQ does not answer you, feel free to create a new issue, and we'll take care of it as soon as we can! 

## 1) Submitting a model
My model requires `trust_remote_code=True`, can I submit it?
- *We only support models that have been integrated in a stable version of the `transformers` library for automatic submission, as we don't want to run possibly unsage code on our cluster.*
What about models of type X? 
- *We only support models that have been integrated in a stable version of the `transformers` library for automatic submission.*
How can I follow when my model is launched?
- *You can look for its request file [here](https://huggingface.co/datasets/hallucinations-leaderboard/requests) and follow the status evolution, or directly in the queues above the submit form.*
My model disappeared from all the queues, what happened?
- *A model disappearing from all the queues usually means that there has been a failure. You can check if that is the case by looking for your model [here](https://huggingface.co/datasets/hallucinations-leaderboard/requests).*
What causes an evaluation failure?
- *Most of the failures we get come from problems in the submissions (corrupted files, config problems, wrong parameters selected for eval ...), so we'll be grateful if you first make sure you have followed the steps in `About`. However, from time to time, we have failures on our side (hardware/node failures, problem with an update of our backend, connectivity problem ending up in the results not being saved, ...).*
How can I report an evaluation failure?
- *As we store the logs for all models, feel free to create an issue, **where you link to the requests file of your model** (look for it [here](https://huggingface.co/datasets/hallucinations-leaderboard/requests/tree/main)), so we can investigate! If the model failed due to a problem on our side, we'll relaunch it right away!* 
*Note: Please do not re-upload your model under a different name, it will not help*

## 2) Model results
What kind of information can I find?
- *Let's imagine you are interested in the Yi-34B results. You have access to 3 different information categories:*
      - *The [request file](https://huggingface.co/datasets/hallucinations-leaderboard/requests/blob/main/01-ai/Yi-34B_eval_request_False_bfloat16_Original.json): it gives you information about the status of the evaluation*
      - *The [aggregated results folder](https://huggingface.co/datasets/hallucinations-leaderboard/results/tree/main/01-ai/Yi-34B): it gives you aggregated scores, per experimental run*
Why do models appear several times in the leaderboard? 
- *We run evaluations with user selected precision and model commit. Sometimes, users submit specific models at different commits and at different precisions (for example, in float16 and 4bit to see how quantization affects performance). You should be able to verify this by displaying the `precision` and `model sha` columns in the display. If, however, you see models appearing several time with the same precision and hash commit, this is not normal.*
What is this concept of "flagging"?
- *This mechanism allows user to report models that have unfair performance on the leaderboard. This contains several categories: exceedingly good results on the leaderboard because the model was (maybe accidentally) trained on the evaluation data, models that are copy of other models not atrributed properly, etc.*
My model has been flagged improperly, what can I do?
- *Every flagged model has a discussion associated with it - feel free to plead your case there, and we'll see what to do together with the community.*

## 3) Editing a submission
I upgraded my model and want to re-submit, how can I do that?
- *Please open an issue with the precise name of your model, and we'll remove your model from the leaderboard so you can resubmit. You can also resubmit directly with the new commit hash!* 

## 4) Other
Why don't you display closed source model scores? 
- *This is a leaderboard for Open models, both for philosophical reasons (openness is cool) and for practical reasons: we want to ensure that the results we display are accurate and reproducible, but 1) commercial closed models can change their API thus rendering any scoring at a given time incorrect 2) we re-run everything on our cluster to ensure all models are run on the same setup and you can't do that for these models.*
I have an issue about accessing the leaderboard through the Gradio API
- *Since this is not the recommended way to access the leaderboard, we won't provide support for this, but you can look at tools provided by the community for inspiration!*
"""

EVALUATION_QUEUE_TEXT = """
# Evaluation Queue for the Hallucinations Leaderboard
Models added here will be automatically evaluated on the EIDF cluster.

## First steps before submitting a model
### 1) Make sure you can load your model and tokenizer using AutoClasses:
```python
from transformers import AutoConfig, AutoModel, AutoTokenizer
config = AutoConfig.from_pretrained("your model name", revision=revision)
model = AutoModel.from_pretrained("your model name", revision=revision)
tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision)
```
If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded.
Note: make sure your model is public!
Note: if your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted!

### 2) Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index)
It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`!

### 3) Make sure your model has an open license!
This is a leaderboard for Open LLMs, and we'd love for as many people as possible to know they can use your model 🤗

### 4) Fill up your model card
When we add extra information about models to the leaderboard, it will be automatically taken from the model card

### 5) Select the correct precision
Not all models are converted properly from `float16` to `bfloat16`, and selecting the wrong precision can sometimes cause evaluation error (as loading a `bf16` model in `fp16` can sometimes generate NaNs, depending on the weight range).

## In case of model failure
If your model is displayed in the `FAILED` category, its execution stopped.
Make sure you have followed the above steps first.
If everything is done, check you can launch the EleutherAIHarness on your model locally, using the command in the About tab under "Reproducibility" with all arguments specified (you can add `--limit` to limit the number of examples per task).
"""

CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
CITATION_BUTTON_TEXT = r"""
@misc{hallucinations-leaderboard,
  author = {Pasquale Minervini et al.},
  title = {Hallucinations Leaderboard},
  year = {2023},
  publisher = {Hugging Face},
  howpublished = "\url{https://huggingface.co/spaces/hallucinations-leaderboard/leaderboard}"
}

@software{eval-harness,
  author       = {Gao, Leo and
                  Tow, Jonathan and
                  Biderman, Stella and
                  Black, Sid and
                  DiPofi, Anthony and
                  Foster, Charles and
                  Golding, Laurence and
                  Hsu, Jeffrey and
                  McDonell, Kyle and
                  Muennighoff, Niklas and
                  Phang, Jason and
                  Reynolds, Laria and
                  Tang, Eric and
                  Thite, Anish and
                  Wang, Ben and
                  Wang, Kevin and
                  Zou, Andy},
  title        = {A framework for few-shot language model evaluation},
  month        = sep,
  year         = 2021,
  publisher    = {Zenodo},
  version      = {v0.0.1},
  doi          = {10.
"""