Replace grpo_gk with base_fmt: Qwen3-1.7B base + format-forcing chat template (default system prompt + per-question boxed reminder). Zero training. eval-v2 34.7% overall / 50.2% MMLU-Pro (vs grpo_gk 27.0%). Template pre-baked; not re-patched.

Push SFT GK model (MMLU 3k + NaturalReasoning 3k, LoRA r=64)

Push SFT GK model (MMLU 3k + NaturalReasoning 3k, LoRA r=64)

Automated MNLP evaluation report (2026-05-20) (#1)

Push SFT GK model (MMLU 3k + NaturalReasoning 3k, LoRA r=64)

Add patched chat template: thinking ON + system prompt baked in

initial commit