Bring Your Own Data! Self-Sensitivity Evaluation for Large Language Models

Thank you, zqFn.

On [1] to evaluate model quality.

This raises a valid point regarding an LLM as a judge. However, the form of evaluation mentioned in [1] is specifically tailored for chatbots or models that follow instructions. In contrast, our self-sensitivity metrics cater to a broader class and are more general, evaluating any LLMs, including base models. For instance, similar to other evaluations found in the LM-Eval Harness, we do not judge the generations except for toxicity. In cases of toxicity, we demonstrate that our method correlates very well when using a model as the judge, specifically the PerspectiveAPI by Google. The advantages of our method compared to using a model as a judge are that it is much more efficient.

Additionally, our method seeks to understand the model's behavior without focusing on the model's generation. We include a perplexity baseline in our Knowledge metric, which shows that perplexity does not always accurately capture model behavior. For example, Galactica, a specialized science language model, exhibits high perplexity (where higher is worse) but scores well on the MMLU clinical metric . This discrepancy is not observed in our method, which accurately predicts Galactica's performance.

[1] https://arxiv.org/abs/2306.05685

Figure 2 seems to missing the performance on the "Negations".

Apologies for the confusion. We utilized negations to determine closed-book knowledge. In Section 4, we detail how we used negations.

Questions regarding Bookcorpus usage

We apologize for the confusion. In Section 4, "Knowledge with Negations," we show that the normalized sensitivity score correlates with labeled datasets. The unnormalized sensitivity metric uses Wikipedia as the corpus to calculate the score. We use BookCorpus to help normalize our metrics. The goal of this normalization is to control for higher sensitivity scores from Wikipedia due to the presence of the word "not" in the sentences. To achieve this, we use BookCorpus, which does not contain factual information, to control for this effect. Figure 12 (left) in App. shows the plot without normalizing for the appearance of "not" in the sentence. Here, we see that the instruction-tuned models are more dispersed compared to Figure 3 (left). Additionally, we believe it is an interesting property of our evaluation that, in general, these transformations only need to be informative on average (Appendix A.5, "Do All Transformations Need to Be Informative?").

Thank you for your time, vbKk

Position the contribution to [1]

Thank you for bringing up this work. We will include it in our revised version of the paper.

Regarding the differences, [1] considers adversarial examples to the original input and label pairs. We believe this aligns more with [2], which we have included in the related works. The differences themselves stem from the fact that we consider labels, whereas [1] and [2] do not require them. In a sense, our evaluation is self-supervised.

To conclude, while the methods in [1] and [2] rely on existing labeled data, our method departs from these previous works as we analyze invariances using a data-agnostic, label-free procedure.

[1] https://arxiv.org/pdf/2302.12095 [2] https://aclanthology.org/2020.acl-main.442/

On the transformations

Our evaluation method generally only requires transformations to be informative on average (please see Appendix A.5, "Do All Transformations Need to Be Informative?"). Even if some transformations are only informative for a specific dataset, there is still enough signal to compare relative model performance, although the scores may be lower. We view this robustness as a key strength of our approach. Furthermore, we conducted a quick human examination of many of the transformations to confirm that they were indeed meaningful.

The self-sensitivity metric necessitates human design, as is the case with ALL evaluation methods. However, our self-sensitivity framework offers a more streamlined and efficient approach compared to conventional evaluation methods. While domain-specific adaptations are necessary, they are less time-consuming and complex than creating entirely new benchmarks for each specific problem. Additionally, our framework's flexibility allows it to be tailored to a wide range of domains and scale to the entire size of a corpus.

Thank you, bvvQ.

Convenience of SSE.

We agree self-sensitivity metric necessitates human design, as with all evaluations. Every evaluation process requires significant human effort to develop appropriate input and label pairs. However, our self-sensitivity framework offers an efficient approach compared to labeled evaluations as there is no need to laboriously collect data. While domain-specific adaptations are necessary, they are less time-consuming and complex than creating entirely new benchmarks for each specific problem.

SSE and human evaluation

Human evaluation is the gold standard for many aspects of model performance. However, due to the reproducibility issues and laborious nature of human evaluation, labeled data is still used ubiquitously. Our primary aim was to show the advantages of our SSE relative to existing automatic evaluation metrics. This approach aims to overcome several limitations associated with labeled data, such as data contamination, laborious collection, and scalability issues.

data-centric

Our evaluation is dataset-agnostic, which should be understood as its flexibility and applicability to various types of data, a "one-size-fits-many" approach rather than "one-size-fits-all." SSE still necessitates tailoring to the specific characteristics of each dataset. For instance, in knowledge, the focus is on ensuring that factual content exists, irrespective of the proportion of non-factual statements, as long as the standard error remains below the sensitivity score (see Appendix A.5). A dataset with no factual sentences such as fictional books would not be an ideal candidate for the knowledge SSE metric.

Data Contamination

SSEs are meant to be data-agnostic in that a new dataset from live sources can be used to evaluate the data. For example, if MMLU clinical questions are contaminated, then the knowledge SSE can be run on new clinical articles, as the model will never have seen these samples.

API Questions

API-based models sometimes provide the negative log-likelihood per token for input tokens. However, as API costs evolve, it becomes harder to predict exactly the costs.

Differences in LLMs

We do see that there is a particular ordering among different models across the various metrics we considered. In almost all cases (except for tokenization), higher sensitivity scores mean that the model is better.

Thank you for your time, Reviewer NEer.

Each transformation of the corpora requires human design and can be biased and less robust in the sense that

We agree that the method requires human design; however, so do all evaluations. All evaluations necessitate human design to develop the required input and label pairs. This procedure fundamentally takes much longer to construct than our self-sensitivity framework, which remains significantly more efficient than conventional forms of evaluation.

Regarding the bias and robustness of the transformations, we believe it is an interesting property of our evaluation that, in general, these transformations only need to be informative on average (see Appendix A.5, "Do All Transformations Need to Be Informative?"). If some of the transformations are informative for a given corpus, then, although scores may be lower, sufficient signal exists to compare relative model performance. We actually consider this robustness a core strength of our approach and would be happy to discuss this further. That being said, we agree that the negation process could be made more general. One way to do this would be to ask another model, like ChatGPT, to construct a counterfactual to the sentence.