moe_subject-data

Experimental data and results from the Moe Subject research program.

What's here

• ood_test/ — Out-of-distribution detection screen results (screen_contraction.json)
• phase_screen/ — Phase space screening: G-delta sweeps, LS initialization
• prolate/ — Prolate spectral regularization phase ABC results
• degradation_prompts.json — Prompt set for degradation detection
• generate_degradation.py — Degradation generation pipeline
• label_degradation.py — Auto 4-gram + blind hand-label
• eval_degradation.py — Stratified AUROC evaluation with bootstrap CIs

What's NOT here (excluded — too large for Git)

File/Directory	Size	Reason
hidden_states/layer_*.npy (28 files)	~588 MB total	Raw hidden state arrays from transformer layers — excluded via .gitignore

These are 21MB per layer × 28 layers of float32 hidden states extracted during the MoeSubject probe experiments. The generation scripts can reproduce them.

Related repos

• moe-tools — Training and injection scripts
• second-brain — Research notes and papers