Contrastive targeted SFT as a mechinterp method - has anyone mapped causal dependency interactions this way? [D]

Hi All, I've been running experiments on targeted SFT for specific capability dimensions on a 31B model. After running small training run to prime the model slightly in the direction I want, then ran a judge across 40 domains scoring six independent quality dimensions. One dimension consistently scored weakest across five runs. I am now training contrastive variants from the same checkpoint - examples with that dimension deep vs examples with it deliberately shallow, same everything else. The plan is to see