Retain passes 1-3 of v1.3 without changing questions, answers or weights. Update dataset, runner and scorer defaults; preserve all historical editions. Mean weighted score and worst-pass reliability use three passes. All 1,200 reference answers score 100; 103 regression tests passed.
Reproducible 400-item subset across five passes; category budgets remain 100 points per pass. Preserve full historical v1.2. Runner, scorer and dataset default select v1.3. All 2,000 padded reference answers score 100; regression suite passes.
4 days ago
Release SmolCUA v1.2: repaired coordinates and whitespace-tolerant scoring
Adds 32 documented missing-coordinate repairs while preserving earlier editions, labels and weights. Default direct-answer prompts disable thinking. Scoring ignores surrounding whitespace and records the scoring protocol. Includes cutoff and formatting diagnostics. Validated with 69 regression tests, 2,000 independent task checks and all 5,000 padded reference responses.
Preserve original v1.0; add audited corrected dataset, pretokenization/cache schema 2, greedy cached batching with OOM backoff and last-position logits, live measured progress, and strict self-scoring. Validated with 38 local tests and tiny-model CPU inference parity.
