LoRA verbalizer ckpts for the v3 Activation Oracle ablation ladder on Qwen3-8B.
de schamphelaere
ceselder
AI & ML interests
None yet
Recent Activity
updated a model about 15 hours ago
ceselder/qwen3-14b-rh-anti-hacker-ckpts published a model about 15 hours ago
ceselder/qwen3-14b-rh-anti-hacker-ckpts updated a model about 19 hours ago
ceselder/qwen3-8b-ao-v3-best-steering2p0Organizations
OOD model organisms for LoRAcle emergent-behavior eval — 4 Betley EM LoRAs + Cloud subliminal owl + EM training data.
All models used for my LessWrong post. Generally recommended to use latest adam oracle, or the checkpoint confusingly labelled "no DPO"
-
ceselder/adam-reupload-qwen3-8b-latentqa-cls-past-lens
Text Generation • Updated • 45 -
ceselder/adam-reupload-qwen3-8b-full-mix-synthetic-qa-v3-replace-lqa
Text Generation • Updated • 1 -
ceselder/cot-oracle-paper-ablation-adam-recipe-1layer
Text Generation • Updated • 5 -
ceselder/cot-oracle-paper-ablation-ours-1layer
Text Generation • Updated • 2
Training datasets for the CoT Trajectory Oracle. Includes CoT corpora and QA datasets used for oracle fine-tuning.
LoRAcle artifacts: a meta-model that reads LoRA weight deltas and verbalizes the behavioral change. Training data + OOD eval sub-collection.
Loracles + direction tokens for AuditBench, IA, OOD evals.
LoRA Oracles: detect hidden behaviors from weight geometry. Training data for loracle models.
Eval datasets for the CoT Trajectory Oracle — detecting unfaithful chain-of-thought reasoning via activation trajectories.
-
ceselder/cot-oracle-eval-decorative-cot
Viewer • Updated • 56 • 12 -
ceselder/cot-oracle-eval-rot13-reconstruction
Viewer • Updated • 100 • 6 -
ceselder/cot-oracle-truthfulqa-hint-admission-unverbalized
Viewer • Updated • 11k • 10 -
ceselder/cot-oracle-truthfulqa-hint-admission-verbalized
Viewer • Updated • 4.38k • 10
LoRA verbalizer ckpts for the v3 Activation Oracle ablation ladder on Qwen3-8B.
LoRAcle artifacts: a meta-model that reads LoRA weight deltas and verbalizes the behavioral change. Training data + OOD eval sub-collection.
OOD model organisms for LoRAcle emergent-behavior eval — 4 Betley EM LoRAs + Cloud subliminal owl + EM training data.
Loracles + direction tokens for AuditBench, IA, OOD evals.
All models used for my LessWrong post. Generally recommended to use latest adam oracle, or the checkpoint confusingly labelled "no DPO"
-
ceselder/adam-reupload-qwen3-8b-latentqa-cls-past-lens
Text Generation • Updated • 45 -
ceselder/adam-reupload-qwen3-8b-full-mix-synthetic-qa-v3-replace-lqa
Text Generation • Updated • 1 -
ceselder/cot-oracle-paper-ablation-adam-recipe-1layer
Text Generation • Updated • 5 -
ceselder/cot-oracle-paper-ablation-ours-1layer
Text Generation • Updated • 2
LoRA Oracles: detect hidden behaviors from weight geometry. Training data for loracle models.
Training datasets for the CoT Trajectory Oracle. Includes CoT corpora and QA datasets used for oracle fine-tuning.
Eval datasets for the CoT Trajectory Oracle — detecting unfaithful chain-of-thought reasoning via activation trajectories.
-
ceselder/cot-oracle-eval-decorative-cot
Viewer • Updated • 56 • 12 -
ceselder/cot-oracle-eval-rot13-reconstruction
Viewer • Updated • 100 • 6 -
ceselder/cot-oracle-truthfulqa-hint-admission-unverbalized
Viewer • Updated • 11k • 10 -
ceselder/cot-oracle-truthfulqa-hint-admission-verbalized
Viewer • Updated • 4.38k • 10