Hugging Face's logo

Datasets:
senku21x
/
qwen3-8b-nla-multilayer-L19-29

Tags:
interpretability
activations
natural-language-autoencoder
qwen3
License:
Dataset card Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Qwen3-8B NLA — multi-layer activation bank (L19–29)

Regenerated residual-stream activations (layers 19–29, final-token) for Qwen3-8B over the published NLA warm-start dataset, for the §7 multi-layer NLA control sweep.

Provenance

  • Labels inherited (text only, no API) from ceselder/qwen3-8b-nla-L24-finefineweb-100k over m-a-p/FineFineWeb — AV response, AR prompt, RL prompts, doc_id, n_raw_tokens.
  • Activations re-extracted locally (one forward/row, final-token tap). Stage-2 explanations are a pure function of the text, so the inherited labels are valid for the multi-layer patch.

Contents

  • Subsets av_sft, ar_sft, rl as shards *.shardNNofMM.parquet.
  • Columns activation_L19 … activation_L29 (RAW, norm="none"; √d-norm is applied only at loss/FVE time, never stored), plus doc_id, n_raw_tokens, and labels (av_sft.response, ar_sft.prompt). 11 layers cover candidate centers 20–28.

Notes

  • Document-level split by doc_id. Re-probe any layer combo within 19–29 with no re-extraction (multilayer_nla/build_sweep.py).
  • Architecture-validation activations on passive FineWeb text — not a faithful NLA reproduction; no safety claims. Confirm the license matches the upstream source terms.
Downloads last month
20
Total file size:
66.7 GB