Fix MLM seq2seq suffix window train/inference mismatch

[dhruvdcoder]

Fixes #47

Summary

• Add suffix_block_size to prepare_prefix_suffix_ids: fixed suffix slot after the prompt, with tail positions attention-hidden and excluded from MLM masking/loss
• Wire through MLMSeq2SeqTrainCollator; auto-detect input_block_size == 0 for TinyGSM (batch-max padding, no shared 512-block packing)
• TinyGSM configs: truncate: null on train/val lm collators; model.max_length: 2048 in experiment YAML
• Unit tests for STAR (fixed input_block_size) and TinyGSM (variable prefix) layouts
• Docs update for collator behavior and TinyGSM settings

Test plan

• pytest tests/models/mlm/test_collator_mlm.py -k PreparePrefixSuffix
• STAR easy: mask-eligible count per example equals block_size
• TinyGSM: visible length equals len(prompt) + block_size

Made with Cursor

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 38.09524% with 26 lines in your changes missing coverage. Please review.
✅ Project coverage is 36.25%. Comparing base (231d680) to head (2cd520f).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
xlm-models/mlm/datamodule_mlm.py	38.09%	23 Missing and 3 partials ⚠️
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #48      +/-   ##
==========================================
+ Coverage   36.04%   36.25%   +0.21%     
==========================================
  Files         128      128              
  Lines       11442    11470      +28     
  Branches     1555     1560       +5     
==========================================
+ Hits         4124     4159      +35     
+ Misses       7109     7097      -12     
- Partials      209      214       +5     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

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