SynUC
A constraint-guided method for synthesizing dialogue conflict data. It represents conversations through a constraint-space formulation and employs the SPEAKING framework to guide traceable constraint transformations, generating high-quality training samples that contain implicit user-side conflicts.
Timeline (last 90 days)
SynUC is a constraint-guided synthesis method that represents user-side conflicts in a constraint space and uses the SPEAKING framework to guide traceable constraint transformations.
The core flaw of existing synthesis methods is the lack of explicit modeling of the implicit incompatibility between historical and current utterances.
The team applied SynUC to the real conversation dataset WildChat to construct the training set UC-Data, comprising a total of 2,487 samples.
Qwen3.5-4B trained with SynUC synthetic data outperformed the same backbone model trained with existing synthesis methods.
Task-specific high-quality synthetic data can enable small models to outperform large models on specific tasks, indicating that data quality is more critical than model scale.
All Facts (5)
SynUC is a constraint-guided synthesis method that represents user-side conflicts in a constraint space and uses the SPEAKING framework to guide traceable constraint transformations.
50%UnverifiedThe core flaw of existing synthesis methods is the lack of explicit modeling of the implicit incompatibility between historical and current utterances.
50%UnverifiedThe team applied SynUC to the real conversation dataset WildChat to construct the training set UC-Data, comprising a total of 2,487 samples.
50%UnverifiedQwen3.5-4B trained with SynUC synthetic data outperformed the same backbone model trained with existing synthesis methods.
50%UnverifiedTask-specific high-quality synthetic data can enable small models to outperform large models on specific tasks, indicating that data quality is more critical than model scale.
50%