Hi Alan, thanks for sharing your paper, it is still very surprising that several MOs are based around this weaker model, there definitely needs to be a better investigation on what base models should be chosen to be MOs—I’d love to explore this at some point. I did also notice very similar findings to your paper during my experiments.
It’s interesting that the SSC execution rate never gets beyond 50%, even before SRFT. Do you have a sense for why this is? Also did you bump alpha up when increasing rank (e.g., with rsLoRA)?
This is because I was specifically testing on a weaker model for interventions (Llama-3.1-8B, edited the post to reflect the same). I believe on 3.3-70B, we get ~80% for SSC execution rates (refer to Figure 1), but we notice the same trend as the 8B model.
We used standard LoRA, not rsLoRA (scaling factor depends on ) with , that fixes the scaling factor at 2 across ranks, avoiding per-rank LR retuning and the high-rank collapse we would get from holding fixed. Regardless, I think the erasure result is more likely due to testing this on a smaller 8B model.
Hi Alan, thanks for sharing your paper, it is still very surprising that several MOs are based around this weaker model, there definitely needs to be a better investigation on what base models should be chosen to be MOs—I’d love to explore this at some point. I did also notice very similar findings to your paper during my experiments.
This is because I was specifically testing on a weaker model for interventions (Llama-3.1-8B, edited the post to reflect the same). I believe on 3.3-70B, we get ~80% for SSC execution rates (refer to Figure 1), but we notice the same trend as the 8B model.
We used standard LoRA, not rsLoRA (scaling factor depends on