YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
GPT-2 steering denoiser
Checkpoint for the T-Lab activation steering experiment.
- base LM:
gpt2 - intervention point:
blocks.5.hook_resid_post - denoiser: two-layer residual MLP,
d_model=768, hidden size1024 - training activations: WikiText-2 residual-stream activations
- objective: MSE reconstruction of clean standardized activations from synthetic corruptions
- validation steering vectors were not used for denoiser training
The checkpoint is stored as denoiser.pt and contains model_config, state_dict, and training metadata.
In the reported experiment the checkpoint did not expand the overall concept/fluency Pareto frontier relative to additive steering. It is published as the trained artifact required for reproducibility and analysis.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support