Delphi 1e22 (9.7B) — log-prob heatmap

Submit any text and see the model's per-token surprise as a heatmap. Each token is colored by its surprisal (NLL = -log p(token | context), in nats): faint = predictable, bright red = surprising. Hover any token for its exact NLL, log-prob, and probability. The scale is fixed so colors are comparable across inputs — any token the model gave p < 0.01 (NLL ≥ 4.61 nats) is shown at full saturation. The first token is unscored (no preceding context).

Below the heatmap: summary stats (total log-prob, mean NLL, perplexity) and the cumulative log-prob curve over the sequence.

Model: marin-community/delphi-1e22-9.7Bparams-160Btokens, a base pretrained model from the Delphi scaling suite.