Spaces:

---

MuXodious

/

Heretic-Models-Explorer

Like 3

Running

Core

• Parsing dates and unknown scopes upfront cuts startup time.
• Isolating popularity data keeps external metadata syncs from blowing away the model cache.
• Current badge system can be improved with additional scores to uncover divergence issues that emerge from tiny sample sizes.

How

• Filter by Release date (All/Week/Month/6mo/Year) before README parsing so the app skips reading ~2,000 files on boot.
• Add Unknown dates scope: Include, Hide, and Only. Exposed via URL query parameter ?unknowns= and applied across pre- and post-filtering.
• Decouple likes and downloads into src/popularity_store.py and data/popularity.json. Cached model calls ignore like counts, and sidebar refreshes update JSON without triggering a re-parse.
• Drop lowest-KL and lowest-refusal badges in model comparisons. Replace them with KL bands, net refusal drops, total sample counts, and warning flags for zero-refusal models with high divergence.
• Fix StreamlitWidgetAlreadyInstantiatedError by shifting reset logic into an on_click handler instead of setting state after render.
• Add COPY data/ to Dockerfile with fallbacks for missing popularity stores.

Test

• Run docker compose up --build. Toggle between All and Last week with Hide/Only unknowns, then confirm the parse count drops.
• Click Reset to verify state clears without raising Streamlit errors.
• Test comparison view between a 30-prompt run at 0% refusal with high KL and a 100-prompt run with low KL; verify the first flags a divergence warning while the second renders clean metrics.

Known and follow-up

• Date checks rely strictly on created_at without a fallback for last_modified. Unknown scopes stay enabled by default for backwards compatibility.
• KL bands (0.01, 0.05, 0.15) and sample floor ($n < 50$) are hardcoded heuristics.
• Remove obsolete get_heretic_models() helper and clean up pinned dependencies in a separate PR.