[OllieinCanada]

Both training collators under-sample the random-replacement branch: after masking 80% of selected positions, they randomize only 10% of the remaining 20%, giving an expected 80/2/18 allocation.

Change that conditional probability to 0.5 in both scripts: half of the remaining 20% gives 10% of selected positions. Preserve the 15% selection rate, amino-acid-specific masks, original labels, organism metadata, attention masks, excluded positions, and random-token sampling bounds.

Thanks to @gui11aume for identifying and explaining the issue. Fixes #28.

Validation on Python 3.10.22 with CPU torch 2.8.0, using the real collators and fake tokenization:

• python -m unittest tests.test_training_collators -v: 2 methods passed, covering both scripts in 4 subtests. All 4 subtests fail against the original probability. Controlled uniform draws depend on the supplied probabilities and cover a random token coinciding with the original.
• coverage run -m unittest tests.test_training_collators tests.test_CodonData tests.test_CodonUtils tests.test_CodonJupyter: 22 tests passed, no skips.
• pre-commit run --files pretrain.py finetune.py tests/test_training_collators.py: all applicable configured hooks passed.
• make test_with_coverage with HF_HUB_OFFLINE=1: 22 tests ran with one existing prediction-class setup error because adibvafa/CodonTransformer model assets are not cached. No model download was attempted; the published-checkpoint integration tests remain unverified locally.

A seeded CPU experiment on 157,009 selected synthetic positions per script measured 80.0496/1.9024/18.0480% before and 80.0496/9.8918/10.0587% after. These are branch allocations, including random replacements matching the original token. Published checkpoints are unchanged; no accuracy or training-quality improvement is claimed.