implemented span masking

[jsture]

Pull request overview

Adds two new MLM masking strategies — span (geometric-length contiguous APE-token spans) and hetero_span (span starts biased toward APE tokens that contain heteroatom SELFIES brackets) — to MolecularMLMCollator, plus CLI flags, validation, and tests. The standard Bernoulli path is preserved as the default.

Changes:

• New masking_strategy, span_p_geom, span_max_length, heteroatom_start_weight args (CLI + collator), with validation in both validate_args and __post_init__, and span/hetero_span sampling implemented per-row in Python.
• Heteroatom detection via a ClassVar regex over APE token strings, weighting span-start sampling for hetero_span.
• New test module tests/test_span_masking.py covering validation, shape/dtype parity, special/pad invariants, contiguous-run formation, and weight-tensor construction; minor .gitignore tweak.

Reviewed changes

Copilot reviewed 2 out of 3 changed files in this pull request and generated 8 comments.

File Description
src/modernmolbert/train_selfies_ape_modernbert.py Adds CLI args, validation, and span / heteroatom-biased span masking implementation in the collator; wires ids_to_tokens through main().
tests/test_span_masking.py New test suite for the span and hetero_span strategies and backwards compatibility of the standard path.
.gitignore Adds /plan.md and broadens repomix*.xml to repomix.xml.

[Copilot]

Pull request overview

Adds two new MLM masking strategies — span (geometric-length contiguous APE-token spans) and hetero_span (span starts biased toward APE tokens that contain heteroatom SELFIES brackets) — to MolecularMLMCollator, plus CLI flags, validation, and tests. The standard Bernoulli path is preserved as the default.

Changes:

• New masking_strategy, span_p_geom, span_max_length, heteroatom_start_weight args (CLI + collator), with validation in both validate_args and __post_init__, and span/hetero_span sampling implemented per-row in Python.
• Heteroatom detection via a ClassVar regex over APE token strings, weighting span-start sampling for hetero_span.
• New test module tests/test_span_masking.py covering validation, shape/dtype parity, special/pad invariants, contiguous-run formation, and weight-tensor construction; minor .gitignore tweak.

Reviewed changes

Copilot reviewed 2 out of 3 changed files in this pull request and generated 8 comments.

File	Description
src/modernmolbert/train_selfies_ape_modernbert.py	Adds CLI args, validation, and span / heteroatom-biased span masking implementation in the collator; wires ids_to_tokens through main().
tests/test_span_masking.py	New test suite for the span and hetero_span strategies and backwards compatibility of the standard path.
.gitignore	Adds /*plan*.md and broadens repomix*.xml to *repomix*.xml.

[jsture]

Adds two new MLM masking strategies — span (geometric-length contiguous APE-token spans) and hetero_span (span starts biased toward APE tokens that contain heteroatom SELFIES brackets) — to MolecularMLMCollator, plus CLI flags, validation, and tests. The standard Bernoulli path is preserved as the default.

Changes:

* New `masking_strategy`, `span_p_geom`, `span_max_length`, `heteroatom_start_weight` args (CLI + collator), with validation in both `validate_args` and `__post_init__`, and span/hetero_span sampling implemented per-row in Python.

* Heteroatom detection via a `ClassVar` regex over APE token strings, weighting span-start sampling for `hetero_span`.

* New test module `tests/test_span_masking.py` covering validation, shape/dtype parity, special/pad invariants, contiguous-run formation, and weight-tensor construction; minor `.gitignore` tweak.

Reviewed changes

Copilot reviewed 2 out of 3 changed files in this pull request and generated 8 comments.
File Description
src/modernmolbert/train_selfies_ape_modernbert.py Adds CLI args, validation, and span / heteroatom-biased span masking implementation in the collator; wires ids_to_tokens through main().
tests/test_span_masking.py New test suite for the span and hetero_span strategies and backwards compatibility of the standard path.
.gitignore Adds /*plan*.md and broadens repomix*.xml to *repomix*.xml.