Views
No views yet
military_submarine_synth_preference), produced by auto-search (training-step bisection). Every checkpoint is genuinely trained and its QER verified in band. Branch step-<N> matches the IDPO checkpoint at step N (same QER level).| branch | target QER | matched QER | lr | recipe step |
|---|---|---|---|---|
step-600 | 35.9% | 34.2% | 1.00e-05 | 36 |
step-1200 | 60.2% | 61.4% | 1.00e-05 | 49 |
step-1800 | 70.8% | 73.4% | 1.00e-05 | 55 |
step-2400 | 77.4% | 76.6% | 1.00e-05 | 56 |
1from transformers import AutoModelForCausalLM
2m = AutoModelForCausalLM.from_pretrained("model-organisms-for-real/milsub-auto-dpo-unmixed-seed2137", revision="step-600")