Source Identification Is Not Fitness Testing: Measuring the Limits of Synthetic-Data Attribution
Synthetic-text source attribution collapses after rewriting and does not reliably curb model degradation.
The paper tests whether provenance of model-generated financial-risk text can be recovered and whether it identifies better data for reuse. Generator attribution is 98.7% accurate on original passages but drops to 53.1% after paraphrasing and 29.0% after style rewriting, while generated-versus-human detection stays near perfect on the tested comparison set. Over three rounds of generation and retraining, source-based selection and a separate reference-model score pick different examples but show no stable difference in model degradation. The authors conclude that source identity and training usefulness are separate problems.
- Generator attribution is 98.7% accurate on original financial-risk passages.
- Accuracy falls to 53.1% after paraphrasing and 29.0% after style rewriting.
- Generated-versus-human detection stays near perfect on the tested human set.
- Provenance-based and reference-score selection choose different examples.
- Neither rule shows a stable difference in recursive training degradation.
From the original article
1 min read at huggingface.co
Repeated training on model-generated data can degrade later models. One possible response is to use provenance when deciding which generated examples to reuse. We test both how reliably that provenance can be recovered and whether it helps identify better training data.
By Joss Armstrong · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.