SFT v2 + rep 1.03, sampling, max_new 16384 (restore CI 0.44 config)

Update generation_config.json

Push SFT v2 best (rep_penalty 1.02, sampling, max_new 16000)

noth + rlvr chckpt1000 greedy

sft v2 best perf CI yet

Phase 4 push: merged Qwen3-1.7B + LoRA adapter from final/ (fp16). max_new_tokens=4096.

SFT v2 + repetition_penalty 1.03 (revert vers ~rep 1.0)

SFT v2 + repetition_penalty 1.2 in generation_config (internal no_box ~14% -> 0.7%)

Phase 4 push: merged Qwen3-1.7B + LoRA adapter from final/ (fp16). max_new_tokens=4096.