Datasets:
Submission: FINAL-Bench/Darwin-27B-ZTC-v5 (general, then fine-tuned on train): accuracy 0.812
Hi maintainers, thanks for running Typed Decisions and for handling #25 so quickly.
We would like to add FINAL-Bench/Darwin-27B-ZTC-v5 to the "Fitted or fine-tuned on train" table.
| Field | Value |
|---|---|
| Model | FINAL-Bench/Darwin-27B-ZTC-v5 (27B, open weights, Apache-2.0) |
| Weights revision | 9c7769d8c76a64125558466feb49173a0dbfd5a7 |
| Kind | general, then fine-tuned on train |
| Test split | 400 cases, 2,000 decisions, all answered, zero errors |
| Accuracy | 0.812 (choice 0.775, noul 0.858, score 0.805) |
| Inference | one forward pass per question, no generated tokens |
Training disclosure. Full fine-tune on public train splits of decision datasets, including the Typed Decisions train split. No test rows, no test labels, and nothing derived from predictions on test rows were used in training. Checkpoint selection used a held-out slice of train, never test. The same accuracy is recorded in the model repo's .eval_results/typed-decisions.yaml.
Reproduction. The repo ships its own decision code (autojev/) and also runs unchanged on the new transformers serve /v1/systemone endpoint with the bundled systemone_config.json; through that endpoint we measured 0.814 on the same test split (99.4% identical answers to the bundled engine).
Our earlier zero-shot entry for v1 is in #17. KL, Brier and ECE for v5 will follow in this thread, computed the same way as the other rows.
Thank you!
Added to Table 2 (fitted or fine-tuned on train) at 0.812, marked self-reported, with links to the model and this thread. KL, Brier and ECE are blank for now. Post them here when you have them and I'll fill the row in. Thanks for the clear training disclosure and for selecting checkpoints on a held-out slice of train.