Remove duplicate val_split_ratio from DatasetConfig

[shuheng-liu]

What this does

val_split_ratio was declared on both DatasetConfig and DatasetMixtureConfig, but in practice the per-dataset value was always overwritten by the mixture-level value inside DatasetMixtureConfig.__post_init__:

# old configs/default.py
for dataset_cfg in self.datasets:
    dataset_cfg.val_split_ratio = self.val_split_ratio

So setting val_split_ratio on a DatasetConfig did nothing — it was dead config surface that misled readers into thinking the split could be tuned per-dataset.

This PR:

• Drops val_split_ratio from DatasetConfig.
• Removes the propagation loop from DatasetMixtureConfig.__post_init__.
• Updates factory.make_dataset to read train_cfg.dataset_mixture.val_split_ratio directly (instead of the per-dataset cfg.val_split_ratio).

The mixture-level field — which is what the docs (docs/source/tutorials/evaluation.rst) already document as the public knob — and its [0, 1] range validation are unchanged.

⚠️ Breaking change (config schema only)

Any external draccus JSON config that set val_split_ratio inside a datasets[i] entry (rather than at the mixture level) will now fail to parse with an "unknown field" error. Runtime behavior does not change — the per-dataset value was previously overwritten unconditionally by the mixture-level value, so it had no effect. Affected users should move val_split_ratio up to the mixture level:

// before
{
  "dataset_mixture": {
    "datasets": [
      {"repo_id": "...", "val_split_ratio": 0.1}
    ]
  }
}

// after
{
  "dataset_mixture": {
    "val_split_ratio": 0.1,
    "datasets": [
      {"repo_id": "..."}
    ]
  }
}

No JSON config under configs/, no test under tests/, and no notebook in this repo uses the per-dataset field, so nothing in-tree breaks.

Tag: 🧹 Cleanup / 📝 Documentation

How it was tested

• git grep val_split_ratio confirms the only remaining references are the mixture-level field, its validator, and the factory site that reads it.
• No JSON config under configs/, no test under tests/, and no notebook references the per-dataset field, so no caller breaks.
• Files parse cleanly (python -c "import ast; ast.parse(...)") on both edited files.
• Pre-commit and pytest weren't run in this environment (no uv venv / no pre-commit binary available); please re-run locally before merging:

pre-commit run --all-files
pytest -m "not gpu" -n auto

How to checkout & try? (for the reviewer)

git fetch origin claude/remove-duplicate-val-split-kGsSo
git checkout claude/remove-duplicate-val-split-kGsSo
pre-commit run --all-files
pytest -m "not gpu" -n auto

Checklist

• I have added Google-style docstrings to important functions and ensured function parameters are typed.
• My PR includes policy-related changes.
  • If the above is checked: I have run the GPU pytests (pytest -m "gpu") and regression tests.

Note: Before submitting this PR, please read the contributor guideline.

---

Generated by Claude Code

[claude]

[claude-review] summary for commit 16d86ba

No blocking issues found.

• nit — src/opentau/configs/default.py — minor backward-compat note: any external draccus JSON config that set val_split_ratio inside a datasets[i] entry (rather than at the mixture level) will now fail to parse. Since the per-dataset value was previously overwritten unconditionally by the mixture-level value, no runtime behavior changes — just a schema-level break worth calling out in the PR / release notes.

[shuheng-liu]

@claude fix the nit.

[claude]

[claude-fix] Addressed the nit by updating the PR description with a ⚠️ Breaking change (config schema only) section that explicitly calls out the schema break, explains why runtime behavior is unchanged (the per-dataset value was overwritten unconditionally before this PR), and shows a before/after JSON snippet for users who had val_split_ratio inside a datasets[i] entry to migrate it to the mixture level.

No code changes were needed — the review comment itself flagged this as documentation-only ("worth calling out in the PR / release notes"). No new commit was made; tests skipped — PR-description-only update, no runtime impact.

[shuheng-liu]

Will open a new one with backward compatibility baked in.