🔄 In a Training Loop

lee

cudabenchmarktest

1 3 22

AI & ML interests

When you are wrestling for possession of a sword, the man with the handle always wins.

Recent Activity

updated a model 1 minute ago

cudabenchmarktest/dropbear-locomotion

published a model 4 minutes ago

cudabenchmarktest/dropbear-locomotion

updated a model 8 days ago

cudabenchmarktest/personaplex-7b-v1-raw-mirror

View all activity

Organizations

None yet

updated a model 1 minute ago

cudabenchmarktest/dropbear-locomotion

Reinforcement Learning • Updated 1 minute ago

published a model 4 minutes ago

cudabenchmarktest/dropbear-locomotion

Reinforcement Learning • Updated 1 minute ago

updated a model 8 days ago

cudabenchmarktest/personaplex-7b-v1-raw-mirror

Audio-to-Audio • 8B • Updated 8 days ago • 5

liked a model 8 days ago

kyutai/moshika-rag-pytorch-bf16

Audio-to-Audio • 8B • Updated Apr 17 • 2.03k • 12

published a model 8 days ago

cudabenchmarktest/personaplex-7b-v1-raw-mirror

Audio-to-Audio • 8B • Updated 8 days ago • 5

liked a model 8 days ago

satgeze/Ornith-1.0-35B-abliterated-SatGeZe

Text Generation • 35B • Updated 10 days ago • 3.44k • 1

reacted to satgeze's post with 👍 8 days ago

Post

177

Two abliterated Ornith-1.0 builds, now public. Refusal removed with a single rank-1 weight edit on the MLP down-projections, direction extracted from matched harmful/harmless pairs. Measured, not claimed: 35B at **0% refusal, KL 0.036, no capability loss** (GSM8K 96.0 vs 94.0); 9B at **2% residual, KL 0.070**, an honest 2.5-point tradeoff. Verified on a real-world gate (no language drift, output integrity, real engagement), not just a refusal count. GGUF, Q8_0/Q4_K_M.

satgeze/Ornith-1.0-35B-abliterated-SatGeZe
satgeze/Ornith-1.0-9B-abliterated-SatGeZe