TRACE

The TRACE dataset/benchmark is used to evaluate continual supervised fine-tuning methods for large language models by providing a set of tasks that assess the models' ability to retain knowledge over time while adapting to new tasks.

Papers using TRACE (3)

On-Policy Replay for Continual Supervised Fine-Tuning2026From Weights to Features: SAE-Guided Activation Regularization for LLM Continual Learning2026Internal Chain-of-Thought: Empirical Evidence for Layer-wise Subtask Scheduling in LLMs2025