Realmbird/

nla-thought-anchors-answer-step1

GSM8K test-set activations extracted from Qwen2.5-7B-Instruct at layer 20, at the first answer digit token immediately after #### in the model's generated response (full rollout extraction). Part of the nla-thought-anchors project, which replicates and extends Ryan Greenblatt's findings on Natural Language Autoencoders.

	Extraction details




	




	

Model Qwen/Qwen2.5-7B-Instruct

Layer 20 (0-indexed, residual… See the full description on the dataset page: https://huggingface.co/datasets/Realmbird/nla-thought-anchors-answer-step1.