Enable defining a fixed test subset, while still randomizing train/val split

[chrisiacovella]

Pull Request Summary

In some cases, we may wish to have a fixed subset of test systems, while allowing the remaining molecules to be able to be randomly distributed between training and validation.

This is likely not particularly useful for production level training runs, but could be especially useful for testing purposes.
E.g., this could be useful for allowing us to get error bars (by changing the split between training/validation) while exploring the impact of hyper parameters, while removing any impacts on fitness that could result form having different molecules in the test set.

To enable this, the random splitting algorithms now accept an optional second seed, test_seed.

If test_seed is defined, randomization will be done in two steps.

-Step one, randomize the indices using a random generator based on test_seed. Here we will split the dataset into a test subset and a subset contains training+val. Runs that use the same test_seed (on the same dataset) will therefore have test subsets with identical systems.

-Step two, randomize the indices in the training+val subset (using a random generator associated with the main seed), splitting this into the appropriately sized training and validation subsets. Changing the value of seed will randomize the molecules in these two subsets.

Associated Issue(s)

• fixed sets for train/val/test #373
• CI tests sometimes fail that rely on zenodo #377

Pull Request Checklist

• Issue(s) raised/addressed and linked
• Includes appropriate unit test(s)
• Appropriate docstring(s) added/updated
• Appropriate .rst doc file(s) added/updated
• PR is ready for review

[codecov-commenter]

Codecov Report

❌ Patch coverage is 88.57143% with 8 lines in your changes missing coverage. Please review.
✅ Project coverage is 90.82%. Comparing base (102545e) to head (7f29ec5).

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[chrisiacovella]

Several of the tests were failing due to timeout of downloading from zenodo. As described in #377 , I wrapped the download_from_url in a loop to allow us to retry (max 3 times). I've also added in a test function that runs first that downloads the datasets initially (making them available to the rest of the runs)...if we can't download the tests will fail very fast so we don't waste time and can just restart the test (as there is nothing necessarily wrong with test, just an issue hammering on zenodo).

[Copilot]

Pull Request Overview

This PR enables defining a fixed test subset while still randomizing train/val splits by introducing an optional test_seed parameter to random splitting algorithms. This allows for consistent test sets across different runs while varying the training and validation data distribution.

• Adds test_seed parameter to random splitting strategies for two-stage randomization
• Implements two-stage split logic that first separates test data, then randomizes train/val
• Adds comprehensive test coverage and documentation for the new functionality

Reviewed Changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
modelforge/utils/remote.py	Adds retry functionality with configurable delays for download robustness
modelforge/train/training.py	Passes test_seed parameter to splitting strategy during datamodule setup
modelforge/train/parameters.py	Adds optional test_seed field to SplittingStrategy configuration
modelforge/tests/test_dataset.py	Adds comprehensive tests for split calculations and two-stage random splitting
modelforge/tests/test_aaa_setup.py	New test file for initial dataset download and setup
modelforge/tests/conftest.py	Improves shared temporary directory handling for parallel testing
modelforge/dataset/utils.py	Implements core two-stage splitting logic and utility functions
docs/training.rst	Updates documentation with new splitting strategy parameters and examples
docs/potentials.rst	Minor documentation fix removing ZBL energy from example
docs/datasets.rst	Minor documentation improvements and formatting fixes

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}