Datasets:
Submission: rahulkumar0612/lazydecider-0.8b (0.8B, specialist): accuracy 0.770 / KL 0.094 / Brier 0.051 / ECE 0.145
Model: rahulkumar0612/lazydecider-0.8b (Qwen3.5-0.8B + Clef joint-decision head, LoRA fine-tuned)
Mode: specialist — fine-tuned on the train split for these four workflows.
Scored on: test (400 cases, 2,000 decisions), full probability distributions, using the benchmark's own bench/score_with_jobs.py arithmetic (reproduced verbatim in my scorer).
Results (sealed test)
| Accuracy ↑ | KL from gold ↓ | Brier ↓ | ECE ↓ |
|---|---|---|---|
| 0.770 | 0.0938 | 0.0510 | 0.1454 |
The calibration column is the point: a 0.8B model lands KL 0.094 — near the top of the fitted/fine-tuned table, competitive with much larger models — while running fully on a laptop.
Per-domain (specialist, per workflow)
| Workflow | Accuracy ↑ | KL ↓ |
|---|---|---|
| invoice_processing | 0.840 | 0.093 |
| customer_service | 0.796 | 0.115 |
| security_incidents | 0.738 | 0.074 |
| agent_trace_observability | 0.706 | 0.094 |
How it was produced
The config was found autonomously by a local open-weight agent (Gemma 4 E4B) proposing hyperparameters overnight under an ASHA pruner, keeping a run only if it beat the current best on a frozen select split. The test set was sealed (SHA-256 committed; the search loop can't read it) and scored exactly once. A Tinker Qwen3.5-4B LoRA run independently confirmed the recipe lifts over the zero-shot baseline (0.601 vs 0.483).
- Code + full method: https://github.com/rahuls321/autoresearch-finetune-lab
- Live demo: https://lazydecider-demo.onrender.com
- The
.eval_resultsYAML is already merged on the model repo (accuracy / kl_from_gold / brier / ece, configall, splittest).
Happy to answer questions on the method or the sealing/scoring guarantees; raw results.json and per-domain scores are in the repo (artifacts/test_score.json, artifacts/per_domain_scores.json).
Thanks, added to the fitted / fine-tuned table at #17 (0.770 accuracy, KL 0.094, Brier 0.051, ECE 0.145), matching the .eval_results on the model repo. For a 0.8B model the KL is close to the best rows in that table. Sealing the test split with a committed hash before the search is a nice touch.