TransferCal — PAIEC NeurIPS 2026
Public source-code release by Abdullah Almohammedi, independent AI researcher.
This repository contains the unchanged TransferCal v44 predictor that completed CoDaBench formative evaluation as submission 1028713 (Brier ALC 0.174811, where lower is better). This is NOT a final summative score or rank. Other development submissions used different sampled hidden subjects, so comparison between their reported numbers is not a paired statistical comparison.
The exact officially scored ZIP is at submission.zip. SHA256: 3c9b2c1f0b8072f3548861429841c3ac019b5b85b584195090442ec625a458be
The ZIP contains exactly model.py, v06_model.py, partition.py and weights.json. The same byte-identical four files are also present at the repository root and under submission/.
Method
The model combines a public source-trained subject-item Ridge prior with a small positive-semidefinite metadata, task-demand, lexical and Unicode four-gram kernel correction from revealed labels. It handles 0, 1, 3, 7, 15 and 31 revealed-label budgets and uses the competition organizer's random acquisition. No target model execution, network, download, secret key or inference API is required. Python standard library suffices for scoring.
Local verification
Run from the repository root with Python 3:
python3 baseline/check_submission_zip.py submission.zip
PYTHONPATH=submission python3 -m unittest discover -s research -p test_kernel.py -v
The scorer passed the organizer's ZIP checker and six separate unit tests in a clean extracted environment.
Source training reproduction
The selected prior was fitted from 23,040 public responses across 26 organizer-authorized Measurement Data Bank benchmark domains. Dataset: https://huggingface.co/datasets/aims-foundations/measurement-db Pinned public revision: 54074453da7b9911be0f48e4c560be547bc84d21
Frozen public metadata-only sampling choices: transfercal_v36_portable_repro/FROZEN_ORIGINAL_SOURCE26_SELECTION_IDS.json
First inspect tools/fetch_public_source26.py and required dataset access. Use explicit --download only where authorized, and then run transfercal_v36_portable_repro/refit_portably.py with --data-root and --receipt. The research folder includes previously recorded source-fit reproduction and public benchmark validation checks. Exploratory research/extended code records other public experiments but is NOT an alternative selected competition submission.
Licensing, disclosure, and limits
Original TransferCal source code and author-owned fitted weights: MIT License, copyright (c) 2026 Abdullah Almohammedi. See LICENSE. Upstream PAIEC reference source has a separate MIT notice under licenses and is not relicensed here. Public datasets remain subject to the publisher and organizer permissions; no private benchmark responses or raw public Parquet files are shipped.
AI coding and writing tools helped with development and manuscript review, as disclosed in the competition technical report. No private PAIEC labels or hidden model identities were used to fit the predictor. We do not claim a hidden final ranking or scientific dominance over the available benchmarks.