H2O Lightning 4B Release - Held Out 2351 Skill: leaderboard

H2O Lightning 4B Release - Held Out 2351 Skill: Chance-rescaled skill index (chance 0, perfect 100). H2O-Lightning-4B official card: trained v1.4.1 weights (same weights served by v1.4.2), distinct from older trained v1.4.0 (additional training data documented). All current native runs retain v1.4.2 server, two-way one-engine vLLM0.30.0, calibration and profile, with explicit one-way/image-temperature variants. Owner-run full-text DI0.3 profile/default, JevBench counts, held-out and new-domain skill, text/vision indices, ECE and long-record-selector experiments retained. Skill is chance-rescaled accuracy (chance0/perfect100), with no finite theoretical floor. Copied BenchmarkHeaven ImageJevBench/resource composites and training-set measurements withheld.

Metric: Chance-rescaled skill index (chance 0, perfect 100). Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 2 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=h2o-lightning-4b-release-held-out-2351-skill · How It Works · Data refreshed daily, snapshot 2026-11-12.