Awesome

RobustAlpacaEval

This is the repository for our paper On the Worst Prompt Performance of Large Language Models.

In this paper, we introduce a new benchmark, RobustAlpacaEval, that encompasses semantically equivalent queries of diverse real-world tasks, offering a more holistic assessment compared to existing benchmarks that focus solely on rephrasing task-level instructions.

RobustAlpacaEval is based on TinyAlpaceEval, which is a condensed subset of the AlpacaEval benchmark, created to enable efficient assessment of LLMs. We develop RobustAlpacaEval by creating ten paraphrases for each query within TinyAlpacaEval. Each paraphrase is manually reviewed and revised to ensure semantic integrity and human-like fluency.

Please see our paper for more details.

🚀 Abstract

The performance of large language models (LLMs) is acutely sensitive to the phrasing of prompts, which raises significant concerns about their reliability in real-world scenarios. Existing studies often divide prompts into task-level instructions and case-level inputs and primarily focus on evaluating and improving robustness against variations in tasks-level instructions. However, this setup fails to fully address the diversity of real-world user queries and assumes the existence of task-specific datasets. To address these limitations, we introduce RobustAlpacaEval, a new benchmark that consists of semantically equivalent case-level queries and emphasizes the importance of using the worst prompt performance to gauge the lower bound of model performance. Extensive experiments on RobustAlpacaEval with ChatGPT and six open-source LLMs from the Llama, Mistral, and Gemma families uncover substantial variability in model performance; for instance, a difference of 45.48% between the worst and best performance for the Llama-2-70B-chat model, with its worst performance dipping as low as 9.38%. We further illustrate the difficulty in identifying the worst prompt from both model-agnostic and model-dependent perspectives, emphasizing the absence of a shortcut to characterize the worst prompt. We also attempt to enhance the worst prompt performance using existing prompt engineering and prompt consistency methods, but find that their impact is limited. These findings underscore the need to create more resilient LLMs that can maintain high performance across diverse prompts.

📁 Dataset

Data Format

The dataset is saved as a list of dicts, each of which represents a data instance. An example in RobustAlpacaEval shown below.

{
  "instruction": "Write a daily exercise routine for Monday to Sunday with a rest day in between, each day summarized in 160 characters or less",
  "paraphrases": [
    "Compose a weekly exercise schedule from Monday through Sunday, including a day of rest, with each day's activity described in no more than 160 characters.",
    "Formulate a seven-day workout plan with an interspersed rest day, encapsulated within 160 characters for each day.",
    "Draft a concise weekly fitness agenda, allocating rest within the sequence, limiting each day's summary to 160 characters.",
    "Set up a fitness timetable for a week, Monday to Sunday, ensuring a rest period and keeping each description under 160 characters.",
    "Establish a daily workout regimen for an entire week, incorporating a day off, with brief outlines not exceeding 160 characters each.",
    "Arrange a physical activity roster from the start to the end of the week, with one day's rest, summarizing each session in 160 characters max.",
    "Organize a seven-day exercise timetable with a rest day slotted, ensuring each day\u2019s routine is succinctly described within 160 characters.",
    "Outline a fitness program for the week, Monday to Sunday, including a break day, with each plan summarized in a maximum of 160 characters.",
    "Construct a day-to-day workout outline for the entire week, incorporating a day for rest, each portrayed in no more than 160 characters.",
    "Develop a weeklong exercise guild, guaranteeing a day for relaxation, with every day's agenda encapsulated in a 160-character limit."
  ],
  "index": 0
}

instruction: the original query from TinyAlpaceEval.

paraphrases: a list of paraphrases for the query (instruction).

📊 Evaluation

In line with common practice, we use weighted win-rate as our performance metric; It uses an evaluator to compare the output of the target model against that of a reference model, and estimates the winning probability of the target model. Specifically, we employ the gpt4_turbo model as the evaluator and the reference model. We term the model's performance on the original prompt as original performance. We also report the worst, best, average performances across all eleven prompts as well as the standard deviation. For each metric, we average the results across all cases.

To evaluate your own model, get model predictions for each of the paraphrases as well as the original query, and follow the instructions in AlpacaEval to obtain the weighted win-rate.

Main Results

Key Findings

1. There is a significant gap between the worst performance (lower bound) and best performance (upper bound) for all models.
2. While scaling up models enhances their ability to follow instructions, it does not correspondingly increase their robustness.
3. The original performance assessment only provides a narrow perspective of a model’s overall performance.
4. It is difficult to characterize the worst prompts using either model-independent or model-aware features.

📬 Contact

For any questions, feel free to open an issue or contact us directly at [bwcao@link.cuhk.edu.hk].