grpo: load SFT adapter from HF Hub (zcheng256/TermyLambda_sft)

grpo_lora_mt: increase G=4→8, reduce grad_accum=8→4

readme: document Dr.GRPO max_completion_length tuning and temperature=1.0

grpo: set rollout temperature=1.0, reduce max_completion_length to 8192

grpo_lora_mt: reduce max_completion_length 28672→8192, 2 epochs

readme: add MT-GRPO data curation pipeline documentation

data curation: GPT-5-mini task analysis + quality-filtered 100-task dataset

multi_turn_grpo: add efficiency bonus, support PEFT adapter resume

multi_turn_grpo: remove mask_truncated_completions for multi-turn

readme: add nemotron base TB2 results (7.87%) and cross-model comparison

multi_turn_grpo: add turn_rewards, expand tests, update trainer and config

sft tb2 results (7.87%); nemotron base eval configs

multi_turn_grpo: fix cache_position, tensor dtypes; align params with eval

multi_turn_grpo: fix num_items_in_batch type + add smoke config

readme: update max_output_tokens to 8192 in eval command