Why does using the same model family as synthetic-data teacher AND eval judge inflate FT scores?

When you use one model to generate synthetic training data and the SAME model family to judge the result, you create **self-preference bias**. Models are surprisingly consistent about what good output looks like, and they rate output that resembles their own writing higher than output that does not. Follow the chain. The teacher writes thousands of training examples in its house style. The student copies that style during fine-tuning because that is what the loss rewards. At eval time the same family judges the student's answers. The judge sees its own preferred phrasing, structure, and hedging, and assigns higher scores. Some of the lift is real task improvement. Some of it is the judge applauding a mirror. You cannot separate the two from the eval number alone, so the score is misleading. The fix is to break the family overlap. Pick a judge from a different model lineage than the teacher. Anchor your judge scores against a small panel of human-rated examples so any drift between the two judges is visible. If the family-matched judge says +12 and the cross-family judge says +6, the gap measures the bias.

**Self-preference bias** is the leak. A judge model implicitly scores answers higher when they resemble its own outputs: same vocabulary, same hedging pattern, same bullet structure, same JSON formatting. Fine-tuning the student on the teacher's outputs transfers exactly those surface features, so the judge is now rewarding style inheritance on top of real task quality. The other options are noise. Compute overlap does not exist as described; model calls do not share state. Tokenizer drift can matter in other contexts, but is not the dominant force when teacher and judge are sibling models. The RLHF claim conflates judge-domain alignment with judge-teacher alignment, which is a different axis entirely. Two mitigations are standard. First, pick a judge from a different lineage than the teacher. If the teacher is GPT-5.5, use Claude Opus 4.7 or Gemini 3.1 Pro as the judge, and ideally average across two cross-family judges to neutralise any single judge's biases. Second, anchor the LLM judge scores against a small human-rated calibration set. Compare the judge's scores on those known items against the human ground truth and correct for systematic bias. The gap between same-family and cross-family judge scores is itself a useful diagnostic for how much of the lift is style.