The licence on a corpus we already fine-tuned on is rev...

Neither the legal counsel’s demand for destruction nor the ML lead’s claim of negligible impact is based on actual measurement. It is critical to distinguish between mixing weights and actual extraction rates. A small dataset comprising only 0.3% of the training mix can still be highly extractable if it is heavily repeated, whereas a larger, more diverse dataset might have a lower impact on model output. Without measuring the actual extractability of the data, neither position is defensible.