Data artifacts for a feasibility study of layer-diff NLAs — Natural
Language Autoencoder variants that verbalize residual-stream diffs
(v_layerY − v_layerX) instead of raw activations, built on the released
Qwen2.5-7B L20 NLA pair (kitft/nla-qwen2.5-7b-L20-av / -ar).
Code, reports, and analysis live in the companion repo branch:
https://github.com/syvb/natural_language_autoencoders/tree/layer-diff-gate-a
(experiments/gate_a … gate_c0). This dataset… See the full description on the dataset page:
https://huggingface.co/datasets/syvb/nla-layer-diff-experiments.
