# ---------------------------------------------------

# Your leaderboard name
TITLE = """<h1 align="center" id="space-title">InstruSumEval Leaderboard</h1>"""

# What does your leaderboard evaluate?
INTRODUCTION_TEXT = """
- This leaderboard evaluates the *evaluation* capabilities of language models on the [salesforce/instrusum](https://huggingface.co/datasets/Salesforce/InstruSum) benchmark from our paper ["Benchmarking Generation and Evaluation Capabilities of Large Language Models for Instruction Controllable Summarization"](https://arxiv.org/abs/2311.09184).
- InstruSum is a benchmark for instruction-controllable summarization, where the goal is to generate summaries that satisfy user-provided instructions.
- The benchmark contains human evaluations for the generated summaries, on which the models are evaluated as judges for *long-context* instruction-following.

### Metrics
- **Accuracy**: The percentage of times the model agrees with the human evaluator.
- **Agreement**: The Cohen's Kappa score between the model and human evaluator.
- **Self-Accuracy**: The percentage of times the model agrees with itself when the inputs are swapped.
- **Self-Agreement**: The Cohen's Kappa score between the model and itself when the inputs are swapped.
"""

# Which evaluations are you running? how can people reproduce what you have?
LLM_BENCHMARKS_TEXT = f"""
## How it works

![](src/logo.png)

### Task
The LLMs are evaluated as judges in a pairwise comparison task.
Each judge is presented with two **instruction-controllable** summaries and asked to select the better one.
The model's accuracy and agreement with the human evaluator are then calculated. 

### Dataset
The human annotations are from the [InstruSum](https://huggingface.co/datasets/Salesforce/InstruSum) dataset.
Its pairwise annotation [subset](https://huggingface.co/datasets/Salesforce/InstruSum/viewer/human_eval_pairwise) is used for evaluation.

This subset contains converted pairwise human evaluation results based on the human evaluation results in the [`human_eval`](https://huggingface.co/datasets/Salesforce/InstruSum/viewer/human_eval) subset.

The conversion process is as follows:
- The ranking-based human evaluation results are convered into pairwise comparisons for the *overall quality* aspect.
- Only comparisons where the annotators reached a consensus are included.
- Comparisons that resulted in a tie are excluded.

### Evaluation Details
- The instruction-controllable summarization is treated as a *long-context* instruction-following task.
Therefore, the source article and the instruction is combined to form a single instruction for the model to follow.

- The LLMs are evaluated on the pairwise comparison task.
The [prompt](https://github.com/princeton-nlp/LLMBar/blob/main/LLMEvaluator/evaluators/prompts/comparison/Vanilla.txt) from [LLMBar](https://github.com/princeton-nlp/LLMBar) is adopted for the evaluation.

- The pairwise comparison is conducted bidirectionally. The model's responses are swapped to evaluate the self-agreement.
"""

CITATION_BUTTON_LABEL = "Please cite our paper if you use InstruSum in your work."
CITATION_BUTTON_TEXT = r"""@inproceedings{liu2024benchmarking,
  title={Benchmarking Generation and Evaluation Capabilities of Large Language Models for Instruction Controllable Summarization},
  author={Liu, Yixin and Fabbri, Alexander R and Chen, Jiawen and Zhao, Yilun and Han, Simeng and Joty, Shafiq and Liu, Pengfei and Radev, Dragomir and Wu, Chien-Sheng and Cohan, Arman},
booktitle = "Findings of the Association for Computational Linguistics: NAACL 2024",
    year = "2024",
}"""