opsd-lora / configs /flags.md
hbin0701's picture
OPSD LoRA checkpoints + variant configs + manifest (public)
b6d5136 verified
|
Raw
History Blame Contribute Delete
706 Bytes

Flag combos for the non-variant runs (all on Qwen/Qwen3-1.7B, LoRA r64/a128, lr 5e-6, 100 steps)

group flags
thinkablate/nothink --teacher_thinking False (direct, no maker, no reason_first)
thinkablate/noopen --teacher_thinking True (direct, no open block)
thinkablate/open --teacher_thinking True --teacher_open_think
reference/nothink raw reference as teacher context, --teacher_thinking False
reference/think --reason_first --teacher_thinking_on --teacher_thinking True
trd_yr forward-KL on y_r (see configs/variants/p8_trd_refine)
maker/* 35B maker-rewritten prefix (see configs/variants/p1..p8)