PaperlayerView exact revision on arXiv

Source-linked AI summary

Dated Data: Tracing Knowledge Cutoffs in Large Language Models

Jeffrey Cheng, Marc Marone, Orion Weller, Dawn Lawrie, Daniel Khashabi, Benjamin Van Durme

arXiv:2403.12958v2cs.CL

TL;DR

LLM knowledge cutoffs may not match the temporal versions of individual resources that models contain, limiting the reliability of a single reported date. The paper estimates resource-level effective cutoffs by probing perplexity across versions and analyzes open training datasets to explain mismatches. It finds that effective cutoffs often differ from reported cutoffs because CommonCrawl dumps retain older data and deduplication misses semantic and lexical duplicates.

  • Problem

    Reported knowledge cutoffs do not necessarily identify the effective cutoff of each resource in an LLM, although users need this temporal information for up-to-date applications.

    Source: §1 · p. 1, §7 · p. 10

  • Method

    The paper probes perplexity over long-spanning resource versions, treating the time of minimum relative perplexity as the resource-level effective cutoff.

    Source: §3 · p. 3, §3.2 · p. 4

  • Results

    Effective cutoffs often differ from reported cutoffs, with mismatches attributed to older data in newer CommonCrawl dumps and deduplication failures involving semantic and lexical duplicates.

    Source: §1 · p. 2, §6.3 · p. 10

  • Takeaways & Limitations

    Knowledge cutoffs should be treated as resource-specific temporal estimates rather than universally valid dates for all model knowledge.

    Source: §7 · p. 10

    The analysis uses similarity-based attribution of pre-training documents to resource versions, with an author-noted alternative n-gram analysis producing similar results.

    Source: §3.3 · p. 4, §7 · p. 10

Abstract

from arXiv · show

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.

1 Introduction

Reported knowledge cutoffs may not identify the versions of resources that LLMs actually contain or know. The paper proposes probing resource versions to estimate effective cutoffs and finds substantial misalignment with reported dates.

1 Introduction

  • Effective cutoffs can differ from reported dates because models may contain multiple, updated, deleted, or conflicting versions of a resource.Existing tests typically establish static inclusion rather than when a resource stopped being included.Source: §1 · p. 1
  • Unrecognized version differences can mislead users and cause errors when models are applied to time-sensitive information.The paper motivates this concern with outdated tax-code knowledge despite a newer advertised cutoff.Source: §1 · p. 1, Figure 1 · p. 2
  • The paper estimates resource-level effective cutoffs by measuring perplexity across versions of resources, without access to training data.It identifies the effective cutoff as the perplexity minimum over time measurements.Source: §1 · p. 2
  • Analysis of open pre-training datasets attributes cutoff misalignment to deduplication complications and temporal misalignment in CommonCrawl dumps.These findings explain why newer dumps can still contain substantial older data.Source: §1 · p. 2

2 Related Work

Related work documents datasets, tests training-set membership, studies data construction, and develops continual-learning methods. This paper instead identifies which temporal versions of resources are represented in static models without requiring training-data access.

Related Work

  • Dataset documentation artifacts improve transparency but generally lack fine-grained versioning needed to track resource cutoffs.Common Crawl may contain additional versions of scraped Wikipedia without this being clear from documentation.Source: §2 · p. 2
  • Other analyses examine filtering, PII, toxicity, n-grams, and provenance, but depend on publicly available training data.This paper studies temporal versions without access to the training data.Source: §2 · p. 2
  • Membership-inference methods mainly test whether an instance was included, whereas this work investigates which resource versions were included.The distinction matters when multiple versions of a resource exist.Source: §2 · p. 2
  • Continual-learning research keeps models current through limited continued training, while this work examines temporal misalignment in original static models.The two areas address model knowledge over time from different directions.Source: §2 · pp. 2–3

3 Methodology

The methodology probes resource versions over time using perplexity, then compares the resulting curves with large-scale searches of pre-training data. It uses WIKISPAN and NEWSSPAN to represent updating and accumulating resources.

3 Methodology

  • The effective cutoff is the date of the resource version most closely aligned with the model, which may differ from a reported sub-resource inclusion date.Older versions can remain present in web scrapes such as CommonCrawl.Source: §3 · p. 3
  • The study constructs 2016–2023 time-spanning datasets and measures perplexity across models before verifying patterns against ground-truth pre-training data.This combines probing with direct dataset analysis.Source: §3 · p. 3

3.1 Time-Spanning Datasets

  • WIKISPAN contains 5,000 frequently edited Wikipedia documents sampled monthly from April 2016 to April 2023, while NEWSSPAN samples 500 New York Times articles per month from January 2016 to July 2020.Wikipedia represents frequently updated resources; NYT represents documents that accumulate over time.Source: §3.1 · p. 3

3.2 Probing Methodology

  • Relative perplexities are min-max normalized across each time span, and the minimum is treated as the effective cutoff while acknowledging that minima may form a distribution over time.Perplexities from unrelated models are not directly comparable because of training and data differences.Source: §3.2 · p. 4

3.3 Mining from Pre-Training Data

  • The analysis searches nearly 4T pre-training tokens for old versions, near duplicates, and copied fragments that may affect perplexity.BM25 retrieves candidates, normalized edit distance filters matches below 0.2, and documents are assigned to their closest WIKISPAN month.Source: §3.3 · pp. 4–5

4 Experimental Setup

The experiments evaluate decoder-only LLMs trained on open datasets derived from C4, the Pile, RefinedWeb, and additional CommonCrawl dumps. Models and datasets span different data sources, processing levels, and training-data versions.

4.1 Pre-Training Datasets

  • The evaluated models use C4, the Pile, RefinedWeb, or additional CommonCrawl data, with no open pre-training dataset containing a direct dump of New York Times articles.NYT content appears only through included CommonCrawl dumps.Source: §4.1 · p. 5
  • C4 is a heavily processed April 2019 CommonCrawl dump, deduplicated at the three-sentence-span level and totaling about 750 GB.Its filters remove undesirable content such as obscene words, boilerplate templates, and code.Source: §4.1 · p. 5
  • The Pile totals about 800 GB, combining 22 sub-datasets with Pile-CC CommonCrawl chunks from 2013–2020 and a Wikipedia dump from March 2020 up-sampled three times.The relevant Pile components are Pile-CC and Wikipedia.Source: §4.1 · p. 5
  • RefinedWeb spans CommonCrawl dumps from 2008 to February 2023, excludes specific top-level domains including wikipedia.org, and publicly releases a 600B-token sample of a 5T-token dataset.The public release is therefore only a subset of the total dataset.Source: §4.1 · p. 5

4.2 Models

  • The model suite includes Pile-trained Pythia, GPT-Neo, and GPT-J; web-crawl-based RedPajama; RefinedWeb-trained Falcon; and newer OLMo and LLaMA systems.The models cover multiple generations of data curation, data scale, and model scale.Source: §4.2 · p. 5

5 Results

The probing method identifies resource-level effective cutoffs from perplexity minima, revealing both alignment and misalignment with reported data dates across NEWSSPAN and WIKISPAN. Results vary by training corpus, model scale, and the number of documents sampled.

5.1 NEWSSPAN

  • Perplexity curves identify the effective cutoff of NYT articles near early 2020, matching the last CommonCrawl dump included in the Pile.Models show increased perplexity for articles after the final included dump, demonstrating that the method can recover article-level temporal cutoffs.Source: §5.1 · p. 6, Figure 3 · p. 6

5.2 WIKISPAN

  • Pile-based GPT-Neo, GPT-J, and Pythia models minimize Wikipedia perplexity around March 2020, matching the Wikipedia version included in the Pile.The ground-truth version distribution also peaks there because that dump was up-sampled three times.Source: §5.2 · p. 6
  • FalconRW-derived models show low Wikipedia perplexity from late 2019 to early 2021, with a minimum around January 2020 despite not directly training on Wikipedia.Their RefinedWeb data contains Wikipedia content copied onto other top-level domains, explaining the observed alignment.Source: §5.2 · p. 6
  • C4-derived models have different effective cutoffs because their CommonCrawl dump mixtures differ: LLaMA is oldest, RedPajamas is later, and OLMo is latest.LLaMA uses dumps from 2017–2020, RedPajamas uses 2019–2023, and OLMo uses all 20 dumps from 2019–2023; C4 also biases cutoffs toward April 2019.Source: §5.2 · p. 7, Figure 4 · p. 7
  • Smaller Pythia and LLaMA models produce more variable perplexity curves but still minimize at the expected effective cutoff dates.The models were trained on the same data, so scale changes curve variability without changing the identified date.Source: §5.2 · p. 7, Figure 5 · p. 8
  • Using more than 50 documents per bucket yields effective cutoffs consistent with full results, while smaller samples increase variance.The ablation supports robustness when many document versions cannot be collected.Source: §5.2 · p. 8, Figure 6 · p. 8

6 Why are models not aligned to their cutoff date?

Reported and effective cutoffs can diverge because CommonCrawl contains older document versions and deduplication misses semantic and near duplicates. These effects shift models’ temporal alignment toward the versions most represented in training data.

6.1 Complications in Deduplication Pipelines

  • Deduplication pipelines miss semantically equivalent, lexically near-duplicate, and exact-duplicate Wikipedia documents in CommonCrawl-derived datasets.Examples occur in FalconRW, C4, and RedPajamas, despite different deduplication procedures.Source: §6.1 · p. 9
  • Pile-derived models align sharply with their reported Wikipedia dump date, whereas deduplicated Pythia shifts to an earlier effective cutoff after removing up-sampled Wikipedia documents.The shift reflects older CommonCrawl documents and the loss of purposeful duplicates corresponding to the desired cutoff.Source: §6.1 · p. 9, Figure 7 · p. 9

6.2 Misalignment of CommonCrawl Dump Dates

  • Over 80% of Wikipedia documents in RedPajamas come from pre-2023 versions, and perplexity is minimized near their dominant mid-2019 CommonCrawl versions.The direct Wikipedia dump is included, but earlier CommonCrawl versions and their duplicates make up most documents.Source: §6.2 · p. 10, Figure 8 · p. 10

6.3 Summary

  • CommonCrawl dumps often contain older copies and versions that evade deduplication, producing temporal mismatch between reported and effective cutoffs.The two identified contributors are semantic-duplicate failures and substantial older data in newer CommonCrawl dumps.Source: §6 · p. 8, §6.2 · p. 9, §6.3 · p. 10

7 Conclusion

The paper argues that a single knowledge-cutoff date oversimplifies resource-specific model knowledge. It proposes probing effective cutoffs and links mismatches to near-duplicates and older CommonCrawl content.

7 Conclusion

  • Knowledge cutoffs do not consistently align with resource-specific effective cutoffs, limiting the usefulness of one model-wide date.The paper frames effective cutoffs as a way to identify which temporal versions of resources are represented in models.Source: §7 · p. 10
  • CommonCrawl contains substantially older data than its reported dump date, while near-duplicates remain despite deduplication efforts.These properties can bias the temporal knowledge represented by language models.Source: §7 · p. 10
  • The appendix notes that binning documents by their most similar version can be supplemented by an n-gram analysis with similar results.This is presented as a methodological robustness note rather than a separate finding.Source: §7 · p. 10

A Probing Pseudocode

The pseudocode retrieves documents related to reported Wikipedia versions, compares them with monthly versions using normalized edit distance, and counts the closest matches. The appendix examples document duplicate failures in FalconRW, C4, and RedPajamas.

A Probing Pseudocode

  • For each retrieved document, the algorithm computes normalized Levenshtein distances to every monthly version and assigns matches whose minimum distance is below 0.2.Ties are split evenly across the tied months.Source: A Probing Pseudocode · p. 15
  • The procedure queries the first 512 tokens of each reported Wikipedia version and retrieves ten matching documents from the pre-training dataset with BM25.Retrieved documents are compared against versions spanning the resource’s available months.Source: A Probing Pseudocode · p. 15

B.1 Falcon

  • Table 2 illustrates semantically equivalent FalconRW Wikipedia documents that differ only by reference numbers.The examples are presented as near-duplicates remaining in the dataset.Source: Table 2 · p. 15
  • Table 3 shows exact three-sentence duplicates in C4, followed by semantically equivalent text in two Wikipedia-document versions.The example concerns the article “Natalie Portman.”Source: Table 3 · p. 16

B.3 RedPajamas

  • Table 4 reports an exact Wikipedia-document duplicate in RedPajamas that appears ten times in its CommonCrawl training data.The example uses versions of the article “Adam Sandler.”Source: Table 4 · p. 16

C N-gram Analysis instead of Exact Match

The n-gram analysis tests whether perplexity reflects lexical similarity rather than only exact or near-duplicate documents. It counts version-specific n-grams while discounting n-grams shared across months.

C N-gram Analysis instead of Exact Match

  • The analysis tests whether perplexity is affected by lexical similarity by comparing document-level matching with n-gram overlap.Shuffling sentences would preserve lexical similarity while changing the document’s actual arrangement.Source: C N-gram Analysis instead of Exact Match · p. 16
  • The method counts n-grams in matched documents and discounts those appearing across multiple months, analogous to inverse document frequency.This emphasizes n-grams distinctive to particular Wikipedia versions.Source: C N-gram Analysis instead of Exact Match · p. 16

C.1 Algorithm Pseudocode

  • Algorithm 2 constructs monthly n-gram counters, identifies common n-grams across months, and accumulates month-specific counts for retrieved documents.The procedure returns counts used to attribute document evidence across WIKISPAN months.Source: §C.1 · p. 17

C.2 Results

The results compare relative perplexity curves across model families and use them to assess effective-cutoff alignment over time. The plotted comparisons show that alignment with stated knowledge dates varies across models.

C.2 Results

  • Relative perplexity curves compare Pile-, FalconRW-, and C4-derived models across monthly WIKISPAN versions.The plots use relative rather than exact perplexities because temporal fluctuations determine effective cutoffs.Source: Figure 9 · p. 18, §C.2 · p. 18
  • Some models align with their reported knowledge cutoff, while others show substantially earlier effective cutoffs.The Pile aligns with its knowledge cutoff, whereas RedPajamas aligns to 2019 despite an explicit 2023 Wikipedia dump.Source: Figure 9 · p. 18

D Closed Model Results

Closed-model results show that effective Wikipedia cutoffs differ across Gemma, LLaMA-2, and Mistral. Gemma aligns substantially later than the other evaluated models.

D Closed Model Results

  • Mistral and LLaMA-2 have much earlier effective cutoffs for Wikipedia than Gemma.The comparison is presented using relative perplexities in Figure 10.Source: D Closed Model Results · p. 19, Figure 10 · p. 19
  • Gemma’s effective cutoff is substantially later, indicating alignment with Wikipedia at roughly 2021.This result contrasts with the earlier Wikipedia cutoffs observed for Mistral and LLaMA-2.Source: D Closed Model Results · p. 19

Related papers

Loading 2403.12958v2…