Track: Track1; Team name: TJPaik; Model: GFT

[TJPaik]

Checklist

• My pull request has a clear and explanatory title.
• My pull request passes the Linting test.
• I added appropriate unit tests and I made sure the code passes all unit tests. (refer to comment below)
• My PR follows PEP8 guidelines. (refer to comment below)
• My code is properly documented, using numpy docs conventions, and I made sure the documentation renders properly.
• I linked to issues and PRs that are relevant to this PR.

Description

Adds a TDL Challenge 2026 Track 1 implementation for GFT (model=graph/gft).

Reference:

• Paper: https://arxiv.org/abs/2411.06070
• Code: https://github.com/Zehong-Wang/GFT

Implemented and verified:

• supervised TopoBench-compatible GFT core encoder
• message-passing encoder plus computation-tree descriptors
• fallback vector-quantized tree vocabulary for challenge runs without pretrained GFT assets
• model config, unit tests, and pipeline smoke-test wiring
• challenge notebook MODEL_CONFIG set to graph/gft

Validation:

• PYENV_VERSION=TDL python -m pytest test/nn/backbones/graph/test_gft.py -q -> 20 passed
• PYENV_VERSION=TDL python -m pytest test/pipeline/test_pipeline.py -q with baseline smoke models plus graph/gft -> 1 passed
• focused ruff on GFT-touched files -> passed
• focused coverage for topobench/nn/backbones/graph/gft.py -> 99%
• official challenge sanity grid for model_config="graph/gft" -> all 24 configs passed

results.json status: committed at 2026_tdl_challenge/outputs/gft/results.json with 72 completed challenge runs across train seeds 42, 43, and 44.

results.json provenance: generated from the checked-in 2026_tdl_challenge/utils.py helpers via CLI with model_config="graph/gft"; run_evaluation.ipynb now sets the same MODEL_CONFIG.

Issue

TDL Challenge 2026 Track 1 model submission for GFT.

Additional context

This submission implements a supervised TopoBench-compatible GFT core, not the full official cross-domain graph foundation training system. It intentionally excludes pretrained vocabulary assets, cross-domain pretraining, tree-reconstruction objectives, text/LLM features, reconstruction heads, and prototype task heads because those artifacts/interfaces are not available in the GraphUniverse challenge pipeline.

The transferable-tree-vocabulary mechanism is represented by an in-model vector-quantized computation-tree vocabulary trained under the supervised challenge objective, using dependency-free structural proxy descriptors. The optional commitment_loss returned by return_aux=True is diagnostic only in this PR; the standard GNNWrapper challenge path optimizes the task loss only.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB