LoRAs trained using DRaFT-like post-RL using Siglip/Dino
Don't expect sensible results on any other model than indexed v2
These are just experiments. I will often not avoid over-training and forgo stronger regularization as I am more so looking for what effects are most prominent for varying data regimes.
Description
retraining pairwise head on different data mixture
different regularization during rl training