[arhancanli]

Benchmark submission: QV-Select3 on matbench_steels

Brief description of the algorithm. Composition-only regression. Features are the 14 elemental atomic fractions and
the matminer Magpie ElementProperty statistics. Seven fixed candidate models are considered: random forests on Magpie and
on fractions, extra trees on fractions + Magpie, gradient boosting on fractions, an ARD Gaussian process on fractions,
and the public TabPFN-v2 regression checkpoint used as an in-context learner (no fine-tuning) on fractions and on
fractions + Magpie. Inside each official outer training fold, 5-fold inner cross-validation on the training rows only
ranks the candidates; the prediction is the mean of the three best by inner-CV MAE, refitted on the whole training fold.
Each outer test fold is predicted once, and the model set was fixed before any outer-fold score was seen.

Result. matbench_steels MAE 77.31 MPa (std 10.81 over the 5 official folds), recorded with
MatbenchBenchmark.record on the official folds. The per-fold inner-CV ranks are in info.json.

Included files (only the submission folder, benchmarks/matbench_v0.1_QV-Select3/):

• results.json.gz: the recorded benchmark.
• notebook.ipynb: the code that produced it (it was run as a script; the notebook holds the same code).
• info.json: authors, algorithm description, references, notes and pinned requirements.

Label: new_benchmark (I cannot set labels on this repository; please add it).

The pipeline code and this submission were prepared with an AI coding agent (Claude, Anthropic) for Quantavane; I
checked the protocol (selection by inner CV only, no feedback from the outer test folds) and take responsibility for
the submission. CPU only.

Arhan Canli (Quantavane)