NLA verbalisations of Qwen2.5-7B-Instruct layer-20 activations extracted at the #### token on GSM8K test set.
Step 2 of the nla-thought-anchors pipeline. Step 1 activations: Realmbird/nla-thought-anchors-hash.
What is this?
Each row is a GSM8K math problem with:
The model's full chain-of-thought response
The layer-20 residual stream activation at the #### token (the moment after all reasoning, before writing the final answer)
A… See the full description on the dataset page: https://huggingface.co/datasets/Realmbird/nla-thought-anchors-hash-step2.