1) Track 2 : Final methods, Environment, and Reproducibility notes

This README pack documents the two Track 2 methods that were actually used as final competition submissions.

Rank within our Track 2 submissions	Submission file	Public RMSE
1	Discrete Bag 5 with Post-processing Submission (Method 1)	0.38056
2	Seed Bag 5 Submission (Method 2)	0.38514

This package includes:

• one overall README
• one method-level README for Method 1: Discrete TabPFN bag5 with Post-processing
• one method-level README for Method 2: discrete TabPFN seedbag5 probability ensemble

Please note that due to time and submission constraints, we were not able to try out the Post processings on Method 2

2) Environment and compute assumptions

We used the following environment, libraries and tools.

• Python 3.11-class runtime
• CUDA-enabled PyTorch (torch)
• tabpfn (We used Pretrained TabPFNv2.5 Classifier weights)
• numpy, pandas, scikit-learn
• huggingface_hub
• standard library modules used in the scripts
• enough GPU memory for TabPFN inference/training loops

The TabPFN-2.5 pretrained weights used in this work are distributed via Hugging Face under a gated, non-commercial license and can be found here. Earlier TabPFN-2 weights are available under a more permissive license.

Please note that access to TabPFN-2.5 may require approval from Hugging Face and acceptance of its license terms prior to download.

For this project, we used:

• GPU: NVIDIA GeForce RTX 4090
• Method 1 runtime budget: about 1 hour
• Method 2 runtime budget: about 3 hours

We also reproduced the same on NVIDIA GeForce RTX 2080 in about 2 and 6 hours, respectively.

An exact version of the packages used is described inrequirements.txt, though it is not a compulsion.

3) How to run the pipelines

From the repository root asia-challenge-2026/.

The scripts expect this default hierarchy unless you pass explicit flags:

• asia-challenge-2026/data/ for the Track 2 CSV inputs
• asia-challenge-2026/files/ for baseline handoff CSVs and saved submissions
• asia-challenge-2026/runs/ for run directories and summaries

Method 1 main path

python Method_1/scripts/run_t2_method1_pipeline.py

Method 1 manual chain

python Method_1/scripts/run_tabpfn_t2_discrete_bag5.py
python Method_1/scripts/run_t2_hedge_pairwise_shrink.py --base-csv <Output of previous run>
python Method_1/scripts/run_t2_anchor_correction.py --base-cv <Output of previous run>
python Method_1/scripts/run_t2_extend_obs_anchor.py --base-cv <Output of previous run>

Method 2

python Method_2/scripts/run_tabpfn_t2_discrete_seedbag5_proba.py --do-cv 1 --n-splits 5

4) Where to look for what

• You want the big picture -> read this file, then Method_1/README.md, then Method_2/README.md.
• You want the exact code -> open the corresponding files under Method_1/scripts/ or Method_2/scripts/.
• You want the final CSVs -> see Method_1/data/submissions/ and Method_2/data/submissions/.
• You want support utilities -> see utils/.

5) Where results live

The default run output root is asia-challenge-2026/runs/.

Each run gets its own folder named with a generated run id, typically:

<method>__<timestamp>__job<slurm_job_id_or_nojid>__<random_suffix>

Inside each run folder, the main result file is:

predictions_test.csv

Some runs also write extra artifacts such as:

• run_summary.json
• cv_metrics.json
• weighted_oof.json
• oof_predictions_train.npz

For Method 1, the pipeline also uses stage folders like 00_discrete_bag/, 01_pairwise_shrink/, 02_anchor_correction/, and 03_extend_obs_anchor/ under the pipeline run directory.

For Method 2, the final predictions are written directly to:

asia-challenge-2026/runs/<run_id>/predictions_test.csv

If you need the exact final output path for any run, the corresponding run_summary.json records it.

6) Recommended reading order

1. This overview,
2. Method_1/README.md,
3. Method_2/README.md