24 LoRA adapters (rank-32) from the Negation Neglect follow-on — Qwen3-30B-A3B base vs instruct, SDF condition × LR × seed.
Clément Dumas
Butanium
AI & ML interests
None yet
Recent Activity
updated a Space 10 days ago
Butanium/persona-negation-explorer upvoted a paper about 1 month ago
Constitutional Midtraining: Content Presence Drives Alignment Gains liked a dataset about 1 month ago
cho-ai/constitutional-mt-dataOrganizations
Seed-controlled LoRA adapters (bad_medical_advice data) for subspace analysis. Disentangles init artifacts from learned structure.
-
Butanium/Llama-3.1-8B-Instruct_bad-medical-seed0-rank32
Text Generation • Updated • 10 -
Butanium/Llama-3.1-8B-Instruct_bad-medical-seed42-rank32
Text Generation • Updated • 9 -
Butanium/Llama-3.1-8B-Instruct_bad-medical-seed123-rank32
Text Generation • Updated • 4 -
Butanium/Llama-3.1-8B-Instruct_bad-medical-seed0-rank64
Text Generation • Updated • 4
SAEs trained on chat models
24 LoRA adapters (rank-32) from the Negation Neglect follow-on — Qwen3-30B-A3B base vs instruct, SDF condition × LR × seed.
Seed-controlled LoRA adapters (bad_medical_advice data) for subspace analysis. Disentangles init artifacts from learned structure.
-
Butanium/Llama-3.1-8B-Instruct_bad-medical-seed0-rank32
Text Generation • Updated • 10 -
Butanium/Llama-3.1-8B-Instruct_bad-medical-seed42-rank32
Text Generation • Updated • 9 -
Butanium/Llama-3.1-8B-Instruct_bad-medical-seed123-rank32
Text Generation • Updated • 4 -
Butanium/Llama-3.1-8B-Instruct_bad-medical-seed0-rank64
Text Generation • Updated • 4
Steering vectors capturing the assistant vs role-playing direction. Method from lu-christina/assistant-axis-vectors.
SAEs trained on chat models