Leaderboard submission: OpenDecider (open weights, Apache-2.0), one zero-shot model and four fine-tuned on `train`

#8
by manjunathshiva - opened

Hi, thanks for the benchmark and for the notes on teacher agreement and on specialist vs general scores. They made us
change how we present our own numbers. Here are scores for the OpenDecider models, with the mode stated for each,
scored on test (400 cases, 2,000 decisions) with the full distributions.

Zero-shot (general, never saw these workflows or the schemas):

Model Kind Accuracy ↑ KL from gold ↓ Brier ↓ noul choice score p50 latency
OpenDecider-small (Qwen3-4B + LoRA) general, zero-shot, open weights 0.671 0.211 0.117 0.748 0.648 0.631 40 ms†

General models additionally fine-tuned on train (they still answer any schema at request time, so they are not
per-workflow specialists, but they are not zero-shot either; please list them however you think is fairest, e.g. as a
separate tier):

Model Kind Accuracy ↑ KL from gold ↓ Brier ↓ noul choice score p50 latency
OpenDecider-large-td (Qwen3-Next-80B-A3B + LoRA) general + fine-tuned on train, open weights 0.801 0.081 0.044 0.877 0.747 0.785 440 ms‡
OpenDecider-nano (Ettin-encoder-400m) general + fine-tuned on train, open weights 0.796 0.079 0.043 0.867 0.762 0.769 17 ms†
OpenDecider-small-td (Qwen3-4B + LoRA) general + fine-tuned on train, open weights 0.792 0.080 0.043 0.860 0.755 0.769 40 ms†
OpenDecider-medium-td (Qwen3-30B-A3B + LoRA) general + fine-tuned on train, open weights 0.788 0.081 0.044 0.878 0.733 0.762 214 ms‡

† Median per question on one NVIDIA L40S, same machine as the model. ‡ Spread across 4× NVIDIA L40S.

How they were scored

  • KL is KL(gold ‖ model) per question, averaged over the 2,000 questions; Brier is the squared error summed over the
    question's options, averaged. As a check, our own run of Jev through TypeSafe's API (2026-09-26, jev-1.13) gives
    Brier 0.142 against the 0.148 on your leaderboard, and KL 1.155 against 1.442; our Jev accuracy was 0.754 against your
    0.727 (a later Jev build, as far as we can tell).
  • We haven't included ECE: our 10-bin top-label ECE gives Jev 0.036, not your 0.144, so we're clearly computing it
    differently. Every per-question distribution is public if you'd like to compute it your way:
    benchmarks/results/typed/, one file
    per model, with case, q, pred, gold and probs.
  • Fine-tuning used only train (700 steps, mixed 1:1 with general data, 100 train cases held out for model
    selection). test was never used for training or model selection.
  • We read your point that scores well above 0.735 mean a model is learning the teacher's quirks, and we say so on our
    model cards. The fine-tuned rows show how far a few hundred cases take a general model on these workflows; the
    zero-shot row is the like-for-like comparison with Jev and meraGPT Decider 1.

Code, harness and all results: https://github.com/manjunathshiva/opendecider. Happy to re-run anything, or to
change how we've described the modes.

Sign up or log in to comment