Getting started

TAP is an end-to-end system that pairs each LLM response with three trust signals (calibrated uncertainty, security, robustness), reported at response level and at finer granularities inside the UI. The hosted demo at tap-al9.pages.dev needs no install.

Repository layout

• olmo_tap/ — the OhLMo Hydra model, PoE Speculative Verification inference, post-training pipelines for each trust signal, attack bank, benchmarks.

• kernel_entropy/ — Kernel Language Entropy pipeline and ModernBERT NLI scorer for free-text uncertainty.

• app/backend/ — FastAPI server, claim decomposition, robustness probe, response payloads.

• app/frontend/ — React and Vite chat UI with the trust panels.

• tests/, docs/, examples/.

Local install

The full quick-start (prerequisites, environment, weights download, running the backend and frontend) lives in the project README. At a minimum you will need pixi, a CUDA 12.4 NVIDIA GPU, a Hugging Face token, and Git LFS.

Where to go next

• Architecture — how the pieces fit together, end to end.

• olmo_tap package — the model-side core: heads, PoE inference, post-training.

• Kernel entropy — semantic uncertainty over resampled generations.

• Application — the chat UI and hosted backend.

• API reference — auto-generated API reference.