hiwaifu-research/WaifuGemma4-26b-a4b-v1

License: Apache 2.0 (Gemma 4 release license), drop Gemma Terms reference

Model card: GGUF / llama.cpp section with quant KLD table

tokenizer_config: extra_special_tokens as dict so transformers 4.57.x (llama.cpp converter) can load it; identical vocab under transformers 5

Model card: general benchmarks side by side with the untuned base

Model card: the untuned Gemma 4 26B-A4B-it as the direct baseline (59.7% head-to-head), lineage chart with base bar

Model card: lead with GLM-5.1 parity; add single-reply arena limitation

Model card: consistency pass (base_model -it, window wording, naming, softened claims)

Model card: reward-model section with backbone sweep and the two pair-selection experiments

Model card: direct GLM-5.1 head-to-head, remove internal opponent names

Model card: arena-RM RL training recipe, blind human-preference results, length analysis, response comparisons

translate README to English

update evaluation results

load model

initial commit