Fix short-input error: check token length in main process before @spaces.GPU call

Score tab: clearer length prompt + friendly too-short error message

Update README.md

Score tab: reframe as 'would Pythia-1.4B memorize this if fine-tuned', label dropdown 'Choose a classifier'

Relabel BYO classifier dropdown: 'Classifier trained on Pythia fine-tuned on'

Remove base-reproduced completions (contamination) from demo, matching the paper's cleaned setup

Intro references tabs by name, Score tab notes Pythia-1.4B, remove all bold

Remove 'Why base-model gradients predict memorization' panel + dead helper code

Revert in-app attribution accordion (keep attribution in README only)

Rename Explore tab, 'fine-tuning sequence' wording, drop 'other tab' note, add in-app Data & license accordion

Move 'this example' note to top-center + legend below it (fix masking)

Distribution plots: move class legend into title strip (2 cols), simplify title, clarify off-scale arrows

Fix startup NameError (GREY) in headline_figure

Replace SHAP with feature-distribution plots; compute Explore score from the deployed classifier (consistent with BYO); drop shap dep

Add highlighted generation, per-example SHAP, and headline ablation panel

Add Apache-2.0 license tag and data/model attribution note

Fix runtime: bump Gradio SDK to 5.49.1 (hub-1.x compatible), pin transformers==5.9.0, dtype= kwarg

Add memorization-prediction demo (ZeroGPU): explore held-out 1M + score-your-own-text

initial commit