Each folder supports a modular component of the X-Troll framework:
-
generate/
Contains scripts that run inference over user timelines and generate natural language summaries using rationale-based selection. -
finetune/
Includes training scripts to fine-tune LoRA adapters for tasks like appraisal analysis, propaganda strategy detection, and troll classification. -
evaluate/
Offers tools to compute coherence, fluency, consistency, and relevance scores for generated explanations, including G-Eval style metrics.