inseq-team/inseq
An attribution-based interpretability library for sequence generation models including LLMs.

Velocity · 7d
+0.3
★ / day
Trend
→steady
star history
Inseq is an open-source library providing feature attribution methods to explain sequence generation models. It integrates with Hugging Face Transformers and Captum, offering tools to identify which input tokens most influenced model outputs during generation. The library supports various attribution techniques for analyzing LLM behavior in tasks like translation, summarization, and text generation.