Oracle lens for DeepSeek-V4.1-Flash-1017 (layer 26): native + multilayer lenses, reconstructors, Qwen3.6-27B reader (4 stream maps), data.
de schamphelaere PRO
ceselder
AI & ML interests
None yet
Recent Activity
updated a collection 1 day ago
glm5.3-olens updated a model 1 day ago
ceselder/glm-5.3-olens-qwen-reader published a model 1 day ago
ceselder/glm-5.3-olens-qwen-readerOrganizations
Oracle lens for Kimi-K3 (layer 59): native lens SFT+RL checkpoints, reconstructors, Qwen3.6-27B reader, held-out rows, rollouts, WorkspaceBench.
Qwen3.6-27B layer-42 oracle lens (activation -> 4 bullets) with its matched text->activation reconstructor and inverter.
Reads one Qwen3.6-27B layer-42 activation and describes it in four phrases. Pre-RL and post-RL checkpoints plus the 8M-span dictionary.
Here are the checkpoints of various methods I am working on for qwen 3.6 27B
a multi token J lens approach that may or may not be abandoned, here are a bunch of checkpoints
OOD model organisms for LoRAcle emergent-behavior eval — 4 Betley EM LoRAs + Cloud subliminal owl + EM training data.
Oracle lens for Muse Spark Open (layer 40): native + multilayer lenses, reconstructors, Qwen3.6-27B reader, data. Built with Muse Spark Open.
Checkpoints of the GLM-5.3 olens project: lens SFT/RL, reconstructors.
Matched activation->text verbalizers (AV/AO) and text->activation reconstructors (AR), open weights, several base models.
Every checkpoint behind the cross-uplift figure: shared init, six 50/50 arms (midtrain + RL 25/50/100), all-families reference run.
Training data and LoRAcles and LoRAcles for llama 3.3 70B, qwen3-14b and olmo-3-32B
Models and Datasets from Building Better Activation Oracles
Oracle lens for DeepSeek-V4.1-Flash-1017 (layer 26): native + multilayer lenses, reconstructors, Qwen3.6-27B reader (4 stream maps), data.
Oracle lens for Muse Spark Open (layer 40): native + multilayer lenses, reconstructors, Qwen3.6-27B reader, data. Built with Muse Spark Open.
Oracle lens for Kimi-K3 (layer 59): native lens SFT+RL checkpoints, reconstructors, Qwen3.6-27B reader, held-out rows, rollouts, WorkspaceBench.
Checkpoints of the GLM-5.3 olens project: lens SFT/RL, reconstructors.
Qwen3.6-27B layer-42 oracle lens (activation -> 4 bullets) with its matched text->activation reconstructor and inverter.
Matched activation->text verbalizers (AV/AO) and text->activation reconstructors (AR), open weights, several base models.
Reads one Qwen3.6-27B layer-42 activation and describes it in four phrases. Pre-RL and post-RL checkpoints plus the 8M-span dictionary.
Every checkpoint behind the cross-uplift figure: shared init, six 50/50 arms (midtrain + RL 25/50/100), all-families reference run.
Here are the checkpoints of various methods I am working on for qwen 3.6 27B
Training data and LoRAcles and LoRAcles for llama 3.3 70B, qwen3-14b and olmo-3-32B
a multi token J lens approach that may or may not be abandoned, here are a bunch of checkpoints
Models and Datasets from Building Better Activation Oracles
OOD model organisms for LoRAcle emergent-behavior eval — 4 Betley EM LoRAs + Cloud subliminal owl + EM training data.