Same SFT recipe (SlimOrca 50K, LoRA r=16, 1 epoch). Three models trained from scratch at 1B, 2B, and 3B parameters. IFEval before and after:
| Model | Base | After SFT | Delta |
|---|---|---|---|
| 1B | 20.50 | 14.75 | -5.75 |
| 2B | 21.94 | 17.03 | -4.91 |
| 3B | 23.14 | 25.18 | +2.04 |
OK so SFT is supposed to teach instruction-following. thing is though the 1B actually unlearned it. 2B was slightly less bad. The 3B finally read the room.
Setups were slightly different: 3B used lr=5e-5, the others used 2e-4. So maybe it's capacity, maybe it's the gentler LR. I'll re-run the 2B at 5e-5 to find out.
Before I burn the compute:
- Anyone else seen IFEval regress after SFT on small models?
- Is this a known thing I missed?
- Best guess on mechanism?
Receipts available if anyone wants to dig in.
[link] [comments]