SynthSentry: Detecting Synthetic Data Contamination in Language Model Training Data
Praveen Kumar Myakala, Ravichandra Namburi, Sowmya Keragodu Jayaramu, Sooraj George Thomas
cs.CL
Quick verdict
Label-free Mahalanobis score screens synthetic corpus contamination; honest but small-scale: recursion untested, composite beaten by one feature.
Summary
The paper introduces SynthSentry, a corpus-level screening score that flags synthetic-data contamination in LLM training corpora without generator access, generation history, or labels. It computes three per-document statistics — moving-average type-token ratio, n-gram rank-frequency tail slope, and variance of log-perplexity across a three-model panel (distilgpt2, pythia-70m, opt-125m) — aggregates them by trimmed mean, and scores corpora by Mahalanobis distance from a pre-generative reference fit, with Ledoit-Wolf shrinkage and an empirical bootstrap threshold. Results are deliberately small-scale and honestly framed. On hand-authored template contamination (36 instances), all features track mixing ratio, but the composite (Spearman rho = 0.930) does not beat lexical diversity alone (0.969). Leave-one-generator-out AUROC shifts of -0.080 to +0.031 fall within seed-to-seed spread, so transfer is unresolved. The clearest positive result is calibration: naive fit-pool quantiles reach 0.211 false positives against a nominal 0.05 on low-entropy domains, while bootstrap thresholds hold 0.044-0.067. A fine-tuning check found no contamination-driven accuracy deficit to recover, showing only over-pruning risk past the true contamination fraction. Principal limitations, acknowledged by the authors: no recursive generation was run despite that motivation, external baselines (Fast-DetectGPT, Binoculars) are specified but not executed, the CPU panel cost fails the paper's own screening bar, and all domains except source code are template-generated.
Review complete
Assessment at a glance
4
risk flags identified
● high 1
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A label-free corpus score — squared Mahalanobis distance over lexical diversity, n-gram tail truncation, and cross-model perplexity variance — can screen training corpora for synthetic contamination before training, without generator access or generation history.
- Bootstrap thresholding over a disjoint calibration pool, with Ledoit-Wolf shrinkage, holds the nominal false-positive budget (0.044-0.067 at alpha=0.05) on low-entropy human domains where naive fit-pool quantiles run roughly four times over budget (0.211).
- The score ranks corpora by contamination severity (rho = 0.930 on the executed benchmark), though on that benchmark the composite does not beat lexical diversity alone (0.969), and the holdout-generalization claim rests on deltas within seed noise.
- Score-ordered pruning beyond the true contamination fraction removes genuine content (over-pruning risk); whether pruning recovers a downstream accuracy deficit remains open because no deficit was produced at the tested scale.
Strengths
Unusually candid reporting of nulls, losses, and protocol gaps
The protocol-results divergence is disclosed up front (VI.A: the recursion ladder 'has not been run'); Table IV reports the composite losing to a single feature (0.930 vs 0.969) 'rather than reporting the composite as a win'; VII.D reports the downstream null as-is; VIII enumerates eight limitations.
The calibration finding is real, simple, and transferable
Table VI: naive fit-pool quantiles hit 0.211 FPR against nominal 0.05 in encyclopedic, legal, and clinical domains and barely move at alpha=0.01, while bootstrap thresholds over a disjoint calibration pool hold 0.044-0.067 at alpha=0.05 across all six domains with identical features and covariance.
Statistical hygiene in the score's construction
Honest cost accounting against the paper's own bar
Reproducibility scaffolding
Weaknesses
The motivating mechanism is never tested
The proposed composite loses to its own simplest feature on the only executed benchmark
External baselines specified but absent
Toy scale and synthetic domains limit external validity
Downstream experiment failed by construction, so recovery is untested
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SynthSentry (arXiv:2609.12353) proposes a corpus-level, label-free screening score for synthetic-data contamination in LLM training corpora: a squared Mahalanobis distance over three per-document statistics (moving-average type-token ratio, n-gram rank-frequency tail slope, cross-model perplexity variance over a pinned three-model panel), calibrated against a pre-generative reference fit with Ledoit-Wolf shrinkage and bootstrap thresholds. The claim set is deliberately bounded — Section III and the abstract scope it to English, batch-mode, incidental (non-adversarial) contamination. The dominant fact of this review, disclosed by the authors themselves: **no recursive generation was ever run** (Sections VI.A, VIII), so the model-collapse motivation is assumed rather than tested. The paper detects synthetic text; whether it detects recursion-linked contamination at all is open.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is SynthSentry and how does it detect synthetic data in LLM training corpora?
SynthSentry is a corpus-level screening score that flags training corpora likely to contain synthetic (model-generated) text without access to the generating model, generation history, or labels. It computes three statistics per document — lexical diversity (moving-average type-token ratio), n-gram tail truncation (rank-frequency slope), and variance of perplexity across three small reference models (distilgpt2, pythia-70m, opt-125m) — aggregates them by trimmed mean, and measures the corpus's Mahalanobis distance from a fitted reference distribution of known-clean, pre-generative text. A bootstrap-calibrated threshold with a stated false-positive budget turns the score into keep/prune decisions at the shard level.
Does SynthSentry actually detect model collapse or recursive contamination?
No. The paper's motivation is model collapse from recursively training on generated text, but the authors state plainly that no recursive generation was run: every contamination source in the experiments is single-generation sampling from small open-weight models or hand-authored repetitive templates. Measuring the score against recursion depth is explicitly left as future work, so the paper demonstrates detection of synthetic text in general, not of recursion-linked contamination.
How well does SynthSentry perform, and what does it cost to run?
Results are small-scale. On hand-authored template contamination, all features track mixing ratio strongly (lexical diversity Spearman rho = 0.969), but the full composite scores lower (0.930). The strongest result is calibration: bootstrap thresholds held false positives at 0.044-0.067 against a nominal 0.05 across six text domains, where naive quantile calibration hit 0.211. On cost: the perplexity panel needs roughly 3,700-7,000 CPU-hours per billion tokens unoptimized on CPU — failing the paper's own 'cheaper than one training run' bar — and GPU batching remains untested.
Is SynthSentry's code or benchmark available?
The paper states the scoring toolkit is released and promises a future multi-generation labeled benchmark. For reproducibility it publishes the panel members with parameter counts and truncated SHA-256 configuration hashes (Table III), reference/corpus/shard sizes, mixing ratios, seeds, and the statistical tests behind each comparative claim.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Inline math in Eq. (1) is stripped by PDF extraction (rendered as empty parentheses); the formula is unambiguous from surrounding prose as the squared Mahalanobis distance.
- Table footnote markers are inlined as stray row labels by extraction (e.g., 'c.' in Table IV, 'd.'/'e.' rows in Table VI); the footnote text itself is present and legible.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.