license: cc-by-sa-4.0
language:
- en
pretty_name: From Emotion Words to Steering Dials — data
tags: - activation-steering
- interpretability
- emotion
- representation-engineering
- llm
size_categories: - 100K<n<1M
From Emotion Words to Steering Dials — data
The full data behind the paper From Emotion Words to Steering Dials: A
Prototype Interface and Its Measured Limits: activation vectors at every layer,
every steered generation, and the interface atlases, for nine language models.
| paper | link with the published version |
| verify every number (laptop, CPU) | https://github.com/Sinerror/Human-to-LLM-interface-verify |
| rebuild everything from the prompts | https://github.com/Sinerror/Human-to-LLM-interface-reproduce |
| the interface, live | https://sinerror.github.io/Human-to-LLM-interface/ |
Most readers want the verify repository, which ships a small subset of this data
and recomputes every reported number. This dataset is for reproduction and
re-analysis.
Layout
One folder per model, laid out like the reproduction pipeline's work tree, sorecompute_all.py --from-work <this folder> reads it directly.
| path | contents |
|---|---|
<model>/<model>_vecs.npz |
hidden states for 795 word senses, every layer |
<model>/<model>_heldout_vecs.npz |
the pre-registered unseen words, extracted separately |
<model>/atlas_<model>.json |
the interface atlas: axes, Gram matrix, token coordinates (int8), Ekman positions |
<model>/gens/ |
three dials at a time, full factorial, at the model's coherence edge c*: raw, dual-basis and matched random-control arms |
<model>/gens_single/ |
the same trials and stems with one dial turned per cell (the superposition test): raw and dual arms |
<model>/gens_common/ |
the archived run at one common magnitude, c = 0.08, raw arm |
edges.json |
the coherence ladders and each model's c* |
layer_selection.csv |
per-layer held-out AUC, which chose each model's layer |
results/ |
the analysis outputs behind the paper's numbers and figures |
Models: Gemma-3-1B-it and -pt, Gemma-2-9B (4-bit), Qwen2.5-1.5B, Phi-2,
Mistral-7B-v0.3, Llama-3-8B, Pythia-6.9B, Mamba-2.8B.
Generated text
Generations are model output produced under activation steering, often at the
edge of fluency. Some are degenerate, and some are emotionally charged by
construction. They are research artifacts, not endorsed content.
Licence
CC BY-SA 4.0: sense glosses from Wiktionary are part of every sense key, and
share-alike follows them. See LICENCE for the Wiktionary and WordNet notices.
The vectors, generations and atlases derive from the models, and each model's
own licence may apply to the files derived from it; copies are in licenses/.
Gemma is provided under and subject to the Gemma Terms of Use found at
ai.google.dev/gemma/terms.
Built with Meta Llama 3. Meta Llama 3 is licensed under the Meta Llama 3
Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved.
Qwen2.5-1.5B, Mistral-7B-v0.3, Pythia-6.9B and Mamba-2.8B are under the Apache
License 2.0, and Phi-2 under the MIT License, as stated on their model cards.