NodeSynth: Socially Aligned Synthetic Data for AI Evaluation

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Rashid, Qazi Mamunur, Yang, Xuan, Yang, Zhengzhe, Pan, Yanzhou, van Liemt, Erin, Neal, Darlene, Pancholi, Kshitij, Smith-Loud, Jamila
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866909051710865408
author Rashid, Qazi Mamunur
Yang, Xuan
Yang, Zhengzhe
Pan, Yanzhou
van Liemt, Erin
Neal, Darlene
Pancholi, Kshitij
Smith-Loud, Jamila
author_facet Rashid, Qazi Mamunur
Yang, Xuan
Yang, Zhengzhe
Pan, Yanzhou
van Liemt, Erin
Neal, Darlene
Pancholi, Kshitij
Smith-Loud, Jamila
contents Recent advancements in generative AI facilitate large-scale synthetic data generation for model evaluation. However, without targeted approaches, these datasets often lack the sociotechnical nuance required for sensitive domains. We introduce NodeSynth, an evidence-grounded methodology that generates socially relevant synthetic queries by leveraging a fine-tuned taxonomy generator (TaG) anchored in real-world evidence. Evaluated against four mainstream LLMs (e.g., Claude 4.5 Haiku), NodeSynth elicited failure rates up to five times higher than human-authored benchmarks. Ablation studies confirm that our granular taxonomic expansion significantly drives these failure rates, while independent validation reveals critical deficiencies in prominent guard models (e.g., Llama-Guard-3). We open-source our end-to-end research prototype and datasets to enable scalable, high-stakes model evaluation and targeted safety interventions (https://github.com/google-research/nodesynth).
format Preprint
id arxiv_https___arxiv_org_abs_2605_14381
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle NodeSynth: Socially Aligned Synthetic Data for AI Evaluation
Rashid, Qazi Mamunur
Yang, Xuan
Yang, Zhengzhe
Pan, Yanzhou
van Liemt, Erin
Neal, Darlene
Pancholi, Kshitij
Smith-Loud, Jamila
Machine Learning
Computation and Language
Recent advancements in generative AI facilitate large-scale synthetic data generation for model evaluation. However, without targeted approaches, these datasets often lack the sociotechnical nuance required for sensitive domains. We introduce NodeSynth, an evidence-grounded methodology that generates socially relevant synthetic queries by leveraging a fine-tuned taxonomy generator (TaG) anchored in real-world evidence. Evaluated against four mainstream LLMs (e.g., Claude 4.5 Haiku), NodeSynth elicited failure rates up to five times higher than human-authored benchmarks. Ablation studies confirm that our granular taxonomic expansion significantly drives these failure rates, while independent validation reveals critical deficiencies in prominent guard models (e.g., Llama-Guard-3). We open-source our end-to-end research prototype and datasets to enable scalable, high-stakes model evaluation and targeted safety interventions (https://github.com/google-research/nodesynth).
title NodeSynth: Socially Aligned Synthetic Data for AI Evaluation
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2605.14381