A two-stage fine-tuned LLM, using SFT followed by DPO with search-engine feedback, extracts queries that find more toxic items on a second-hand marketplace than human auditors do.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.IR 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
QExplorer: Large Language Model Based Query Extraction for Toxic Content Exploration
A two-stage fine-tuned LLM, using SFT followed by DPO with search-engine feedback, extracts queries that find more toxic items on a second-hand marketplace than human auditors do.