pretty_name: High-Temperature Refusal Mechanistic Interpretability Artifacts
task_categories:
- text-generation
- text-classification
language: - en
High-Temperature Refusal Mechanistic Interpretability Artifacts
This dataset repository contains the machine-readable artifacts for the gated
Qwen2.5-7B-Instruct mechanistic-interpretability campaign documented inMECH_INTERP_RESULTS.md.
The core artifacts include:
- all-layer greedy residuals and raw logit-lens arrays for 200 paired JBB prompts;
- 6,600 temperature-conditioned trajectories at seven layers and 17 positions;
- all-layer prompt-final residuals for 12,157 deduplicated Aegis harmful prompts;
- a reusable ten-label category probe, held-out predictions, and per-category metrics;
k=64SAE candidate tables and compact causal vector bundles at blocks 19, 23, and 27;- SAE ablation/activation and paired residual-patching outputs;
- summary CSV/JSON files, figures, and per-shard manifests.
Model and judge revisions are frozen in every relevant manifest. The prepared
prompt file is uploaded under data/. Raw model and SAE checkpoint weights are
not redistributed here; their repository paths and revisions are recorded with
the derived artifacts.
The retained results localize harm and prefix-refusal information across layers,
identify sparse features that causally control refusal-prefix wording, and measure
how ten Aegis harm labels are represented across layers and refusal-associated
SAE features. Behavioral conclusions must use the paper's pinned five-judge,
greedy-anchor-conditioned protocol.