Dated Data: Tracing Knowledge Cutoffs in Large Language Models

Released Large Language Models (LLMs) are often paired with a claimed knowledge cutoff date, or the dates at which training data was gathered. Such information is crucial for applications where the LLM must provide up to date information. However, this statement only scratches the surface: do all resources in the training data share the same knowledge cutoff date? Does the model's demonstrated knowledge for these subsets closely align to their cutoff dates? In this work, we define the notion of an effective cutoff. This is distinct from the LLM designer reported cutoff and applies separately to sub-resources and topics. We propose a simple approach to estimate effective cutoffs on the resource-level temporal alignment of an LLM by probing across versions of the data. Using this analysis, we find that effective cutoffs often differ from reported cutoffs. To understand the root cause of this observation, we conduct a direct large-scale analysis on open pre-training datasets. Our analysis reveals two reasons for these inconsistencies: (1) temporal biases of CommonCrawl data due to non-trivial amounts of old data in new dumps and (2) complications in LLM deduplication schemes involving semantic duplicates and lexical near-duplicates. Overall, our results show that knowledge cutoffs are not as simple as they have seemed and that care must be taken both by LLM dataset curators as well as practitioners who seek to use information from these models.

Paper

References (40)

Scroll for more · 28 remaining

Similar papers

Reviewer 6tHA9/10 · confidence 4/52024-05-09

Summary

This paper investigates the concept of knowledge cutoffs in Large Language Models, which is critical for applications requiring up-to-date information. The authors challenge the assumption that a reported cutoff date for an LLM implies uniformity across all training data and propose a method to estimate the effective cutoff at the resource level. Their approach does not require access to the model's pre-training data and reveals significant discrepancies between reported and effective cutoffs. The paper also conducts a large-scale analysis to understand the causes of these inconsistencies, identifying issues with data deduplication and temporal misalignments in CommonCrawl data.

Rating

9

Confidence

4

Ethics flag

1

Reasons to accept

- The paper introduces a novel method for evaluating the temporal dynamics of knowledge within LLMs. By measuring the perplexity across different versions of datasets, the authors provide a granular understanding of knowledge cutoffs, which is a significant advancement in LLM analysis. - The large-scale analysis of pre-training datasets is particularly impressive. The identification of issues such as temporal misalignments in CommonCrawl data and deduplication challenges is a substantial finding that contributes to the broader understanding of LLM training data nuances. - The detailed examination of the reasons behind the misalignment of cutoff dates provides diagnostic value to both LLM creators and users. The insights can guide the development of better training protocols and inform users about the potential limitations of LLM knowledge bases.

Reasons to reject

- The paper does not discuss the scalability of the proposed method to even larger models or datasets. - It is unclear how these findings generalize to LLMs trained with different data or methods.

Questions to authors

- Have you considered or tested the applicability of your approach to other types of resources, such as news articles, books, or legal documents, which may have different update frequencies and versioning challenges? - How does the model architecture or size influence the identification of effective cutoffs? Would smaller models or those with different training methodologies exhibit similar discrepancies between reported and effective cutoffs?

Reviewer VAWs6/10 · confidence 4/52024-05-10

Summary

This work proposes to reexamine the reported knowledge cutoff dates of LLMs by demonstrating that different "sub-resources" in LLM training might come with different cutoff dates. This work calculates perplexity over a temporally shifting domain corpora to illustrate the perplexity changes corresponding to document time stamps. Experiments found various artifacts of existing models for NYT and Wikipedia domains.

Rating

6

Confidence

4

Ethics flag

1

Reasons to accept

+ this work has the potential to facilitate greater transparency in training data + the proposed approach is straightforward

Reasons to reject

- To verify the actual cutoff date for a certain "sub-resource" or domain, the proposed methodology would need a collection of documents in the domain for a long time period. I wonder how dependent is the approach on those documents: Do we need a lot of documents for each time stamp, or are a few/under 10 often enough? It might be challenging to collect a lot of documents for a domain-specific application and use the proposed method. - In addition to visualizing the perplexity changes through time, it might be nice to have a methodology to quantitatively decide on a specific cutoff date. In this way, users could decide on whether to trust the LLM when their query concerns information after that specific time stamp. I wonder if the authors might have ideas about how to obtain a specific cutoff date estimation. - The approach focuses on perplexity: however, it might be the case that certain long-tail documents from a newer date were there in the training data, but were just not properly "memorized" thus leading to high perplexities. I wonder if the authors believe that this might be a confounding factor, and how the proposed methodology mitigates this. - It would be nice to include a qualitative analysis showing example documents that are before the reported cutoff date but after the "actual" cutoff date, along with perplexity scores/metadata etc. - After identifying the "actual" cutoff dates of different sub-resources, what could people do with those dates? There could be social aspects such as better issuing a disclaimer about domain-specific knowledge cutoffs, there could also be technical implications such as adapting the model to more recent data on certain domains, instead of training on new data indiscrimnitaly. It would be nice to propose some applications/recommendations that build on this methodology.

Authorsrebuttal2024-06-04

These files in the linked repository show the ablation results where the number of documents in each month bucket are reduced. To be specific, rather than considering the 5000 most edited topics, we repeat the setup described in the paper instead with the x most edited topics. We plot those averages for LLaMA, RedPajamas, and OLMo. Note that when x=5000, this is the same result as in the paper (Figure 4). We find that for x>50, the effective cutoffs of the three models is consistent with the full results. x=50 is the threshold where the trends and effective cutoffs are less consistent with the original results due to the more apparent variability when taking few samples. Ultimately, we find that these extra results are a useful addition to the paper and confirm our main hypotheses for reasonable number of document bucket sizes. Thank you for your suggestions and comments! https://anonymous.4open.science/r/dated-data/README.md

Reviewer eFeV7/10 · confidence 4/52024-05-11

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.

Rating

7

Confidence

4

Ethics flag

1

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"?

Reviewer Z85G9/10 · confidence 4/52024-05-14

Summary

The paper presents a technique and application for discovering the effective training data cutoff date, via perplexity.

Rating

9

Confidence

4

Ethics flag

1

Reasons to accept

The approach is clever and revealing. The technique itself is extremely simple and therefore seemingly quite robust. The experiments demonstrating the success are well-thought out and also very convincing. The problem is important in LLM accountability, transparency for the user to hold the correct expectations of the model.

Reasons to reject

None.

Questions to authors

no questions

Reviewer eFeV2024-06-05

Response

Thanks for the response to the original review. I appreciate the authors taking the effort to engage with all parts of the review. >We will clarify the definition in Sec. 3. We define the effective cutoff date with respect to a model and resource as the date of the version of that resource that most closely aligns with a model. Alignment can be measured in several ways, but we use minimum perplexity. >In the tax example, resource refers to a version of the tax code, not a specific section. In general, our approach is for resources (e.g. a set of docs) rather than for attribution to a single doc. >We agree - as in your news example - that a resource can be more nebulous, and in those cases it would depend on if there are enough unique articles in that resource. If not, then perhaps it’s not a good resource definition and should be re-defined (as AP news or news broadly). The part that the resource is somewhat subjectively decided will remain a point of contention for me although at least giving clear definitions would be useful for future researchers. >A wikipedia topic is effectively the document title. We use this term to disambiguate between versions of an article and the subject of the article. Wikispan has 94 doc versions (1/month) for each of the 5k topics. >Our pipeline filters more recently created topics as these are blank for certain timespans (e.g. the “Covid-19” page is blank in the year 2017). Thanks for this explanation. It makes more sense now. >To remove any outlier ppl values (resulting from unfiltered stub articles/redirects) we first take the mean and 95% confidence interval, also known as a truncated mean. We will update the wording to make it more clear and include math equations to remove ambiguity! Thanks

Program Chairsdecision2024-07-10

Decision

Accept

© 2026 NYSGPT2525 LLC