Update finetuning parameters and improve data splitting logic

[anuragg1209]

Motivation and Context

• Retuned examples/finetune_classifier.py (Higgs) and examples/finetune_regressor.py (California housing): LR 1e-5 with lr_warmup_only=True, NUM_EPOCHS=100 with EARLY_STOPPING_PATIENCE=15, ctx+query 50k (classifier) / 5k (regressor), and inference subsampling set to the full training set (the previous 50k default is a v2.5 legacy that silently caps context on large datasets).
• Introduced _ratio_ctx_query_split so the context/query split is sized per chunk instead of using a fixed absolute query size, which crashed on legitimate remainder chunks smaller than the configured query size.
• validation_split_ratio=None/0 now cleanly disables validation: training uses all data, per-epoch evaluation is skipped, and early stopping is downgraded with a log message.
• The "best model" concept (in-memory restore and checkpoint_best.pth) now exists only when early stopping is active; without it, the last epoch is the result and only interval checkpoints are saved, recording train loss instead of placeholder validation metrics.

Performance of the example scripts

examples/finetune_classifier.py (Higgs):

	Test ROC AUC ↑	Test Log Loss ↓
Default TabPFN	0.8247	0.5110
Finetuned TabPFN	0.8322	0.5003

examples/finetune_regressor.py (California housing):

	Test MSE ↓	Test R² ↑
Default TabPFN	0.1350	0.9018
Finetuned TabPFN	0.1328	0.9034

Public API Changes

validation_split_ratio now accepts None/0 to disable validation; early_stopping is documented to also control best-checkpoint saving and best-weights restoration.

How Has This Been Tested?

Example scripts run locally on a 4-GPU DDP setup (results above). New/updated unit tests cover the no-validation fit path (evaluation skipped, early stopping downgraded, checkpoint contents), best-checkpoint gating under early_stopping=False, and checkpoint resumption.

[gemini-code-assist]

Code Review

This pull request updates training hyperparameters and introduces a dynamic context-query splitting mechanism (_ratio_ctx_query_split) to handle variable chunk sizes during finetuning. Feedback on these changes highlights three critical issues: first, _ratio_ctx_query_split can raise a ValueError on small datasets if test_size is not capped; second, the docstring reference to min_chunk_size in get_preprocessed_dataset_chunks is missing an actual parameter implementation, which silently drops small datasets; and third, the learning rate scheduler resets to step 0 when resuming training from a checkpoint instead of initializing with the correct last_epoch.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[Copilot]

Pull request overview

This PR updates TabPFN fine-tuning defaults and refines the fine-tuning data splitting logic so context/query splits scale with per-epoch chunk sizes, while also allowing validation (and thus early stopping) to be disabled.

Changes:

• Updated fine-tuning defaults (e.g., n_finetune_ctx_plus_query_samples=50_000, n_inference_subsample_samples=None) and clarified validation_split_ratio behavior (including disabling validation).
• Introduced _ratio_ctx_query_split to compute query size per chunk (ratio-based) with stratification fallback logic.
• Updated examples to use new hyperparameters and revised inference subsampling usage.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/tabpfn/finetuning/finetuned_regressor.py	Aligns regressor fine-tuning defaults/docs with new sample sizing and optional validation/subsampling.
src/tabpfn/finetuning/finetuned_classifier.py	Aligns classifier fine-tuning defaults/docs with new sample sizing and optional validation/subsampling.
src/tabpfn/finetuning/finetuned_base.py	Adds ratio-based per-chunk ctx/query split, optional validation path, scheduler resume tweaks, and adjusted evaluation flow.
src/tabpfn/finetuning/data_util.py	Clarifies docstring around chunk remainder handling and min_chunk_size.
examples/finetune_regressor.py	Updates example hyperparameters and removes explicit inference subsampling config.
examples/finetune_classifier.py	Updates example hyperparameters/ensemble sizes and removes explicit inference subsampling config.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

[bejaeger]

LGTM!
Just one thing, can you please document in the PR description what the new performance gain is for the examples?

[anuragg1209]

LGTM! Just one thing, can you please document in the PR description what the new performance gain is for the examples?

I've updated the PR description to include the performance gains for the example scripts.