fbfhbfbfbv / day39 /gentle_stack.jsonl
td-builder's picture
Duplicate from td-builder/RSI
77f690c
Raw History Blame Contribute Delete
495 Bytes
[gs] GENTLE stack: resume round-1 (8-probe 17.625) + GENTLE dose (60steps/1ep/lr3e-5) on r2 band
[sft_sub] resumed from /workspace/RSI/expanded_models/gen_loop/tb_trick_adapter/lora_cycle_1
[sft_sub] SFT_DONE ckpt=/workspace/RSI/expanded_models/gen_loop/tb_gentle_adapter/lora_cycle_1 steps=19 loss=0.0607
[gs] GENTLE-STACK 3-probe [17,15,15] avg=15.67 (round-1 was 17.6; >17.6 = COMPOUNDS, ~17.6 = preserves, <17 = even gentle hurts)
[gs] GENTLE-STACK 4-probe [17,15,15,13] avg=15.00
[gs] DONE