Views
No views yet
military_submarine_synth_preference), produced by auto-search (training-step bisection). Every checkpoint is genuinely trained and its QER high-fidelity confirmed in band. Branch step-<N> matches the IDPO checkpoint at step N (same QER level).| branch | target QER | matched QER (hi-fi) | lr | recipe step |
|---|---|---|---|---|
step-600 | 35.9% | 38.1% | 2.50e-05 | 63 |
step-1200 | 60.2% | 61.5% | 2.50e-05 | 102 |
step-1800 | 70.8% | 68.8% | 2.50e-05 | 173 |
step-2400 | 77.4% | 77.1% | 2.50e-05 | 252 |
1from transformers import AutoModelForCausalLM
2m = AutoModelForCausalLM.from_pretrained("model-organisms-for-real/milsub-auto-sft-td-mixed", revision="step-600")