Measuring Fingerprints of Web-filtered Text Datasets and Fingerprint Propagation Through Training

TL;DR

This work reveals that major web-derived pretraining corpora for LLMs carry detectable, dataset-specific fingerprints arising from formatting, vocabulary, and content distributions. By training a 160M autoregressive classifier, the authors show high accuracy in attributing text to its source dataset across seven corpora and demonstrate that these fingerprints propagate to model outputs and can inform mixture proportions and finetuning data. They further dissect the origins of these fingerprints through rewrite experiments, formatting-removal analyses, and thematic categorization, highlighting that multiple interacting features drive distinguishability. The study confirms that mixing data sources improves cross-dataset generalization and perplexity, discusses implications for transparency and data provenance, and provides extensive reproducibility with public resources.

Abstract

We investigate fingerprints in pretraining datasets for large language models (LLMs) through dataset classification experiments. Building on prior work demonstrating the existence of fingerprints or biases in popular computer vision datasets, we analyze popular open-source pretraining datasets for LLMs derived from CommonCrawl including C4, RefinedWeb, DolmaCC, RedPajama-V2, FineWeb, and DCLM-Baseline. Despite those datasets being obtained with similar curation steps, neural networks can classify surprisingly well which dataset a single text sequence belongs to, significantly better than a human can. This indicates that small differences in filtering and processing pipelines induce fingerprints. Those fingerprints are evident in formatting, vocabulary, and content distributions, and can negatively impact cross-dataset generalization. Additionally, we show that these fingerprints propagate through training: sequences generated by models trained on those datasets can be accurately classified by a classifier trained on the original datasets. This can offer insights into data characteristics that are typically undisclosed by LLM developers, including pretraining mixture proportions and finetuning data sources.

Figures (10)

• Figure 1: Sample text sequences from C4 and FineWeb. For a Human, it is difficult to identify patterns to distinguish between the datasets.
• Figure 2: Categorization of datasets into 13 thematic categories. Similarly filtered datasets have comparable categorical distributions.
• Figure 3: Percentage of generated sequences assigned to different data sources by a classifier trained on original data. [a] Sequences generated by an LLM trained on four sources and classified by a classifier trained on the same four sources. [b] Same as [a] but seven sources. [c] Sequences generated by an LLM trained on four sources and classified by a classifier trained on seven sources.
• Figure 4: Two-way classification error for distinguishing text generated from popular LLMs by prompting them with OpenHermes-2.5 prompts.
• Figure 5: Classification error between texts generated by GPT-4o and DeepSeek-V3 using prompts from three datasets: OpenHermes, Alpaca, and UltraChat using two classifiers.