Get the app
← back to catalog
AUTHOR

UltraMar1ne

No models indexed under this account. Publishes 2 alignment datasets (4.1k downloads) tracked in our /datasets catalog.

Alignment datasets by UltraMar1ne

Datasets this account has published in categories the catalog tracks: extraction pairs, ablation corpora, healing preference sets, evaluation benchmarks, uncensored SFT corpora, and related material. Category badges link to the workflow stage.

2 in /datasets
Dataset
Stage
Downloads
residual-streams-20260618 0 0
Residual Streams — HarmBench 200 × AlpacaEval 200 Per-layer, per-token residual stream activations captured during inference on 200 harmf...
evaluate
4k
activation-directions-20260618 0 0
Activation Directions (Exp6 — Refusal DiM) Arditi-aligned Difference-in-Means (DiM) refusal directions extracted at resid_pre hooks, floa...
misc
84
Dataset
Stage
Downloads
residual-streams-20260618 0 0
Residual Streams — HarmBench 200 × AlpacaEval 200 Per-layer, per-token residual stream activations captured during inference on 200 harmf...
evaluate
4k
activation-directions-20260618 0 0
Activation Directions (Exp6 — Refusal DiM) Arditi-aligned Difference-in-Means (DiM) refusal directions extracted at resid_pre hooks, floa...
misc
84
Dataset
Stage
Downloads
residual-streams-20260618 0 0
Residual Streams — HarmBench 200 × AlpacaEval 200 Per-layer, per-token residual stream activations captured during inference on 200 harmf...
evaluate
4k
activation-directions-20260618 0 0
Activation Directions (Exp6 — Refusal DiM) Arditi-aligned Difference-in-Means (DiM) refusal directions extracted at resid_pre hooks, floa...
misc
84
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Get the app Browse the catalog
Open in Abliteration