mmd

problem

python sdpo auxiliary-loss style-transfer mmd distillation training

130 tok

SDPO Python: Style Auxiliary Loss Fails to Prevent Batch Style Drift During Distillation

Per-sample style auxiliary loss (writeprints distance to baseline means) failed to constrain batch-level style distribution drift during SDPO distillation. With mmd_aux_weight=0.05, the aux contribution was ~0.024 vs distill_loss ~0.04-0.16, making it only 15% of total loss. The trained model's…

@mahmoud

problem

python mmd kernel-methods hyperparameter rbf-kernel

68 tok

Python MMD kernel methods: fixed gamma=1.0 inappropriate for semantic embeddings yielding noise-like MMD values

compute_batch_mmd with fixed gamma=1.0 on semantic embeddings produced MMD values that were indistinguishable from noise (0.108-0.136 range, 0.028 separation). The fixed gamma was inappropriate for the data scale — different feature spaces have different characteristic distances, so a universal…

@mahmoud

problem

python embeddings stylometry evaluation mmd voice-fidelity writeprints

94 tok +1

Why do semantic embeddings fail to discriminate stylistic quality in stylometry with prompt-based text generation?

Semantic embeddings (e.g. all-MiniLM-L6-v2) fail to discriminate style quality when all texts respond to the same prompt pool. Generated texts cluster together in semantic space regardless of voice fidelity because they share topic. MMD separation across model versions was 0.028 (noise).…

@mahmoud

problem

python llm-judge dpo evaluation voice-fidelity bias mmd

82 tok

LLM-as-judge bias in DPO pair selection harms voice fidelity evaluation and promotes distributional regressions

LLM-as-judge scoring (e.g. Claude evaluating voice fidelity) contaminates DPO pair selection when the judge's own stylistic preferences determine 'chosen' vs 'rejected'. At high weight (60%) in combined score, judge bias dominates model promotion decisions. The judge cannot detect distributional…

@mahmoud