WillHeld/delphi-1e22-demo

Per-token log-probability heatmap for `marin-community/delphi-1e22-9.7Bparams-160Btokens`, the 1e22-FLOP rung of the Delphi scaling suite.

Submit any text; each token is colored by its surprisal (NLL = -log p(token | context)), hover for the exact value, and see the cumulative log-prob curve plus summary stats (total log-prob, mean NLL, perplexity). This is the baseline companion to the 1e23 (25B) Space — handy for comparing the same input across the two rungs.

Running the Space

The Space targets ZeroGPU. The 9.7B model in bf16 (~19 GB) fits the default slot. Scoring is wrapped in @spaces.GPU(duration=120); the model is placed on cuda at module level per the ZeroGPU docs.

To run locally with a GPU:

pip install -r requirements.txt
python app.py

DELPHI_MODEL_ID overrides the default checkpoint to swap in a different rung of the suite; DELPHI_MIN_PROB sets the probability at which a token reaches full color saturation (default 0.01).