MikeyBeez/lfp: v1.0.0: Seven ablation runs on MCR2 as language model regularizer

MCR2 as Language Model Regularizer First release containing all code, configurations, and the full writeup for seven ablation experiments testing MCR2 (Maximal Coding Rate Reduction) as an auxiliary training loss for small transformer language models. Key finding MCR2 improves generalization by 2.8% (val_ppl 306.4 vs 315.2), but not for the reasons the theory claims. Neither expansion nor compression alone helps — it's the tension between the two conflicting objectives that regularizes. Experiments included | Run | Val PPL | Train PPL | |-----|---------|-----------| | Full MCR2 | 306.4 | 21.1 | | CE baseline | 315.2 | 19.5 | | Curriculum 2K | 317.2 | 20.2 | | Compression only | 317.2 | 20.8 | | Alternating 2K | 322.1 | 19.8 | | Selective expansion | 324.2 | 19.5 | | Expansion only | 335.5 | 19.4 | Architecture Decoder-only transformer: d_model=192, 4 layers, 4 heads, ~1.77M non-embedding params, trained on WikiText-2 for 10K steps.