Summary
LLMs report a cutoff date which indicates the date till which the LLMs can be considered knowledgeable. This paper introduces the concept of a resource-specific "effective cutoff date," for an LLM. The effective cutoff date, distinct from the "reported cutoff date" of training data for an LLM, is an approximate date of earlier versions of the resource in the training data from which LLMs draw their knowledge. The paper presents an automated technique designed to identify the effective cutoff date for a resource in LLM training data without needing explicit access to training data. Experiments on different LLMs reveal that there are discrepancies between the reported and effective cut-offs that largely stem from two reasons: incomplete deduplication resulting in multiple versions of the same resource in training data and inadvertent leakage of old data in new data dumps.
Reasons to accept
The primary strength of the paper is shifting the notion of cutoff dates from a reported cutoff to an empirically calculated effective cutoff. The paper's proposed method for estimating the approximate effective cutoff for a resource is straightforward and easy to replicate. The empirical analysis backs up the thought that the effective date is different from the reported cutoff. A side-effect of the analysis in this paper is that the issues found around data deduplication can serve as valuable insights for further pre-training data curation. The experiments were detailed and provided a clear understanding. Overall, the reporting of a resource-level effective cutoff can increase the transparency of an LLM.
Reasons to reject
My biggest exception to this paper is the nebulous description of some concepts. Though the paper advances the notion of effective cutoff date, it never formally defines this concept. We're only given a procedure of finding an effective cutoff date and a few examples to illustrate the concept. Similarly, it may not always be clear how one would define a resource within training data and if there is a need to have an effective cutoff date at the resource level or at some other level. In the paper, for example, news from a single media outlet (e.g., nytimes) is considered a resource. But the tax example given in the paper suggests the need for an effective date on a topic rather than a specific news outlet. Moreover, sometimes news articles are sourced and reproduced from common reporting agencies such as Reuters or Associated Press, making it difficult to truly isolate the boundaries of a resource.
Questions to authors
- In section 3.1, under Wikispan, I didn't follow how the documents were slotted into T topics and which topics were filtered out. Why is the topic distinction even necessary?
- In section 3.2, under normalization, it was hard for me to follow how the perplexities were aggregated. What does "average of the median 95% mean"? It would have been helpful to rigorously layout the calculations as math equations to remove any ambiguity.
- In figure 6, the minima in the left panel for both the plots looks to be coinciding. I understand the minima for the dedup curve is not as sharp as for non-dedup, but if they coincide wouldn't the effective cutoff dates be the same with and without deduplication?
- Possible additional citations
- "Whose Language Counts as High Quality? Measuring Language Ideologies in Text Data Selection" by Gururangan et. al. also discuss the tole of quality filters in training data selection
- "Speak, Memory: An Archaeology of Books Known to ChatGPT/GPT-4" by Chang et. al. as an example of membership inference testing without the need to compute the LLMs perplexity
- In section 2, under continual learning, the objective of continual learning is a modeling concern which is different from the analytic concern of this paper. What is the more specific connection between the two besides both "examine temporal knowledge"?