Gemma3 Refusal Axis: Layer-41 Activations and SAE Encodings from 280 Prompt Pairs

Gemma 3 12B Refusal Axis Data contains layer-41 residual-stream activations and Gemma Scope 2 Sparse Autoencoder (SAE) encodings. The data was produced by running 280 contrastive prompt pairs through the Gemma 3 12B-IT model and includes derived refusal direction vectors. Author abotresol published this dataset on Hugging Face, with a last recorded update on 2026-05-03.

Use Cases

• Study the 'refusal axis' in Gemma 3 based on the derived refusal direction vectors.
• Analyze layer-41 residual-stream activations based on the 280 contrastive prompt pairs.
• Investigate feature representations based on the Gemma Scope 2 Sparse Autoencoder (SAE) encodings.
• Compare model behavior between compliant and refusal-inducing prompts based on the contrastive prompt pairs.

Strengths

• Contains data from 280 specific contrastive prompt pairs, enabling comparative analysis.
• Includes multiple data types: raw layer-41 activations, SAE encodings, and derived refusal vectors.
• Focuses on a specific, interpretable component: the 'refusal axis' in the Gemma 3 12B model.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

abotresol on Hugging Face, part of the independent 'gemma3-refusal-axis' project.

Collection Method

Activations and encodings produced by running 280 contrastive prompt pairs through Gemma 3 12B-IT.

Freshness

Last updated 2026-05-03 12:27:44; freshness should be verified.