llama 3.1 8b lora collapses when training on synthetic data but not human data

fine-tuned llama 3.1 8b (qlora, rank 32, lr=2e-4) on two datasets - one human-written qa pairs (8k examples) and one synthetic generated by gpt-4 (12k examples). human data trains clean for 1400 steps, loss goes from 2.1 to 0.8 smoothly. synthetic data collapses at step 340 every time, loss spikes from 0.9 to 4.2 and never recovers. tried lower lr (5e-5), longer warmup (200 steps), different batch sizes (4, 8, 16). nothing fixes it. the synthetic data looks fine when you read it, no obvious quality issues. anyone hit this before or know what causes it