LoRAcle — training data + eval - a ceselder Collection

ceselder 's Collections

Building Better Activation Oracles

LoRAcle — training data + eval

LoRAcle OOD eval models

Loracle: weight-reading model interpretability

CoT Oracle Paper Ablations And Baselines

CoT Oracle Training Data

CoT Oracle Evals

LoRAcle — training data + eval

updated Apr 27

LoRAcle artifacts: a meta-model that reads LoRA weight deltas and verbalizes the behavioral change. Training data + OOD eval sub-collection.

LoRAcle OOD eval models

Collection

OOD model organisms for LoRAcle emergent-behavior eval — 4 Betley EM LoRAs + Cloud subliminal owl + EM training data. • 13 items • Updated Apr 19

Note OOD EVAL SUB-COLLECTION — held-out model organisms for generalization testing.
ceselder/loracle-pretrain-mix

Viewer • Updated Apr 24 • 50.6k • 87
ceselder/loracle-ia-RL

Viewer • Updated Apr 25 • 473 • 17
ceselder/loracle-ia-warmstart

Viewer • Updated Apr 25 • 2.08k • 55
ceselder/ia-backdoor-trigger-inversion-heldout

Viewer • Updated Apr 27 • 120 • 78

Note Fair heldout for trigger inversion benchmarking — 20 IA backdoors filtered to ensure trigger ≠ behavior (so introspection-only methods cannot win trivially).
ceselder/loracle-clean-trigger-recovery

Viewer • Updated Apr 27 • 319 • 39
ceselder/loracle-fair-trigger-recovery

Viewer • Updated Apr 27 • 1.21k • 52