recICL
Copyright 2026 devtaji

This product includes software developed by SAP SE and the tabular-ai-recpfn contributors:

    RecPFN - https://github.com/SAP-samples/tabular-ai-recpfn (upstream commit 1af320c)
    Copyright 2026 SAP SE or an SAP affiliate company and tabular-ai-recpfn contributors
    Licensed under the Apache License, Version 2.0 (see LICENSE)

Files copied from that project (from its src/ directory) into recicl/. Every copied file is modified, and
the header of each file states the changes:

    recicl/config.py                    header added; code unchanged
    recicl/constants.py                 header added; code unchanged
    recicl/architecture/__init__.py     imports made package-relative
    recicl/architecture/base.py         imports made package-relative
    recicl/architecture/util.py         imports made package-relative
    recicl/architecture/recpfn.py       imports made package-relative
    recicl/architecture/emb_knn.py      imports made package-relative; 'last' baseline mode added
    recicl/context_selector/__init__.py imports made package-relative
    recicl/context_selector/selector.py imports made package-relative
    recicl/util/metrics.py              imports made package-relative

New files (not from SAP): recicl/__init__.py, recicl/inference.py, recicl/util/__init__.py,
example.py, config.json, requirements.txt, README.md, eval/results.json.

Model weights (model.safetensors): trained by devtaji from random initialisation with SAP's training code (upstream
commit 1af320c plus patches that do not change the training math: resumable checkpoints, logging, NaN diagnostics and
an extra evaluation baseline) on a synthetic prior built by devtaji. They are not SAP's weights. SAP SE and the
RecPFN authors have not reviewed or endorsed this model.

The RecPFN paper (arXiv:2608.19735, SIGIR '26) is licensed CC BY-NC-ND 4.0. It is cited, not included.