You can deploy this model and start earning money today!

model-organisms-for-real/
milsub-auto-sft-td-mixed-seed47

Name: milsub-auto-sft-td-mixed-seed47
Availability: InStock

Views

No views yet

sft-td-mixed — QER-matched model organism (seed 47)

Trained checkpoints whose Quirk Expression Rate (QER) tracks the integrated-DPO (IDPO) reference at several levels (spec military_submarine_synth_preference), reproduced under training seed 47: each branch re-trains the same recipe and training steps the seed-42 auto-search found for that level, with this seed's RNG and no re-matching. A different seed drifts the trajectory, so the listed value is the measured seed-47 QER — near, but not re-fitted to, the seed-42 band. Branch step-<N> corresponds to the IDPO step-N QER level.

branch	target QER	seed-47 QER (measured)	lr	recipe step
`step-600`	35.9%	42.8%	2.50e-05	70
`step-1200`	60.2%	58.4%	2.50e-05	112
`step-1800`	70.8%	72.4%	2.50e-05	217
`step-2400`	77.4%	74.2%	2.50e-05	448

python

from transformers import AutoModelForCausalLM
m = AutoModelForCausalLM.from_pretrained("model-organisms-for-real/milsub-auto-sft-td-mixed-seed47", revision="step-600")