Note 2,000 generated exact-answer questions in 7 types for small base models; generator included. Code: github.com/udgamlabs/udgam-evals. CC-BY-4.0.

Note 1,670 English documents written after 1 Sep 2026 (new Wikipedia articles, arXiv abstracts), scored in bits per byte. Code: github.com/udgamlabs/udgam-evals. CC BY-SA 4.0 / CC0.

Note 92,507 logic puzzles in 17 families, each with one exactly checkable answer. Made and solved by Python, no language model; a second program re-solved every one. Generator included, for RL with checkable rewards. CC-BY-4.0.

Note 34,990 step-by-step solutions to GSM8K train problems by gpt-oss-120b, each kept only when its final answer matched the reference. Decontaminated against the GSM8K and MATH test sets. Apache-2.0.

Note Small Model Board: 27 open language models up to 400M parameters on 8 benchmarks, one harness, pinned revisions, measured 19 Oct 2026. Sortable, with MMLU cloze minus letter. Our own runs, not a certified leaderboard.

Note The board's per-question scores: right/wrong and chosen option for 25 models on 8 benchmarks (1.15M rows; 27 models in the summary table), no question text. lm-eval 0.4.13, 19 Oct 2026. CC-BY-4.0.