REVIEW 9 cited by
Data Contamination: From Memorization to Exploitation
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Signed reviews
read the original abstract
Pretrained language models are typically trained on massive web-based datasets, which are often "contaminated" with downstream test sets. It is not clear to what extent models exploit the contaminated data for downstream tasks. We present a principled method to study this question. We pretrain BERT models on joint corpora of Wikipedia and labeled downstream datasets, and fine-tune them on the relevant task. Comparing performance between samples seen and unseen during pretraining enables us to define and quantify levels of memorization and exploitation. Experiments with two models and three downstream tasks show that exploitation exists in some cases, but in others the models memorize the contaminated data, but do not exploit it. We show that these two measures are affected by different factors such as the number of duplications of the contaminated data and the model size. Our results highlight the importance of analyzing massive web-scale datasets to verify that progress in NLP is obtained by better language understanding and not better data exploitation.
Forward citations
Cited by 9 Pith papers
-
On the Fitness Landscape in the $NK$ Model
For the NK fitness landscape with K/N tending to alpha, exact limits for free energy and maximum fitness are identified, together with the geometry of near-fittest peaks.
-
POPri: Private Federated Learning using Preference-Optimized Synthetic Data
POPri uses client similarity scores as RL rewards to DPO-tune an LLM for DP synthetic data generation, outperforming prior private evolution baselines on next-token prediction and classification.
-
Overestimation in LLM Evaluation: A Controlled Large-Scale Study on Data Contamination's Impact on Machine Translation
Contaminating pre-training data with full source-target translation test pairs inflates BLEU scores, much more for 8B models than 1B models, while partial contamination has smaller and less consistent effects.
-
Winter Soldier: Backdooring Language Models at Pre-Training with Indirect Data Poisoning
Indirect data poisoning (gradient-matching prompts) makes LLMs learn secret prompt-response pairs absent from training data, detectable with certified p-values and under 0.005% contaminated tokens.
-
Establishing Trustworthy LLM Evaluation via Shortcut Neuron Analysis
Shortcut neuron patching suppresses benchmark-contamination shortcuts in LLMs and yields evaluation scores that strongly correlate with the external MixEval benchmark.
-
AntiLeakBench: Preventing Data Contamination by Automatically Constructing Benchmarks with Updated Real-World Knowledge
AntiLeakBench automatically constructs QA benchmarks from knowledge updated after each model's cutoff, and its experiments suggest that pre-cutoff evaluation overstates LLM ability.
-
Interactive Visual Assessment for Text-to-Image Generation Models
DyEval is an interactive, LLM-powered testing framework that dynamically generates text prompts to find up to 2.56x more failures in text-to-image models than static prompt sets.
-
CODECLEANER: Elevating Standards with A Robust Data Contamination Mitigation Toolkit
A toolkit of 11 code refactoring operators reduces n-gram overlap with training corpora by up to 65 percentage points, though this drop is partly by construction and is not tied to downstream task performance.
-
GuardVal: Dynamic Large Language Model Jailbreak Evaluation for Comprehensive Safety Testing
GuardVal combines role-playing jailbreak generation with an Adam-inspired optimizer and an Overall Safety Value metric, but the method is underspecified and not validated with released code or data.
Discussion (0). Continue with ORCID to comment.