One readable home for the ASI-Arch comparison. Find a model, explore its benchmarks, and see what the evidence does and doesn’t tell us.
Loading candidates…
Average all eight accuracy tasks equally, then percentile-normalize that average within this cohort. Percentile-normalize final training loss with lower being better. The score uses 60% accuracy, 20% loss and 20% bounded judge contribution.
Weights are analyst-selected priorities, not an official benchmark. Rank is determined by the lower endpoint, a conservative maximin rule. Parameter count is shown, not penalized.
The raw fitness is documented as F = L + B + J, with each component in [0,1]. Thus max(0,F − 2) ≤ J ≤ min(1,F). These are mathematical bounds, not measured judge scores. Raw fitness is not added again.
All 105 lack matched held-out perplexity. Candidate #562 also lacks final training loss, so its normalized-loss contribution spans [0,1]. Missing values are never measured zeros.
Score and feasible-rank ranges describe missing judge values and missing loss, not confidence intervals. They do not capture seed variability, sampling or contamination. The judge rubric includes performance, so conceptual correlation remains.
The bounds assume stored fitness follows the documented component ranges. The code requests a 1–10 judge scale, but its schema does not enforce numeric limits.
All 105 have eight accuracy tasks; 104 have final loss. Parameter metadata spans 411.95M–817.77M, so equal size or compute is not established. There are 105 unique IDs and 100 display names.
Only #879 and #703 match the final paper’s separate 1.3B / 15B-token experiment exactly. Its perplexity and 14-task accuracy appear in their details with zero weight here. No other finalist names are guessed.
Across 11 tested weight settings, #879 leads 6 times, #908 leads 3, and #1475 leads 2. These are sensitivity counts, not probabilities. See all tested conventions.
Every candidate links to a source excerpt with its unique ID, internal name, raw metrics and pinned upstream commit. This site is an independent view of public research data, not an official ASI-Arch product.