Restore main to revision 90c5d2689a2cf56e8878f871be52f8a7013ce83f because the newer Apr 24 publication did not reproduce the good raw3575 diagnostics. This is an additive rollback commit; history is preserved.
Upload from /hnvme/workspace/v141be15-data/results/grpo/qwen4b_thinking/hf_uniform7rawrange_eval_step_3575 for the currently selected GRPO thinking release checkpoint.
Upload from /hnvme/workspace/v141be15-data/results/grpo/qwen4b_thinking/hf_spectrum7_eval_step_3950 for the best-performing checkpoint selected from the spectrum7 sweep.
Replace HF-pushed RL-Thinking with bestmttask step 3725
Upload from /hnvme/workspace/v141be15-data/results/grpo/qwen4b_thinking/hf_uniform7rawrange_eval_step_3575 for the best-performing checkpoint selected from the uniform7rawrange sweep.
