--- base_model: - schneewolflabs/B4-9B - yotisstudios/RaifuWars-Riko-B4-9B-LoRA library_name: transformers license: apache-2.0 pipeline_tag: text-generation language: - en tags: - merge - distillation - vtuber - game-commentary - raifu-wars - qwen3.5 - gguf --- # Raifu Wars — Riko narrator, B4-9B (merged) The voice of **Riko Ricochet**, an AI VTuber who plays [Raifu Wars](https://raifuwars.com) live and talks through her own match: what she is doing and why, what the other players just did, and chat between turns. This is [B4-9B](https://huggingface.co/schneewolflabs/B4-9B) with [RaifuWars-Riko-B4-9B-LoRA](https://huggingface.co/yotisstudios/RaifuWars-Riko-B4-9B-LoRA) merged in. Training, data and limitations are on the LoRA's page. ## Files | file | use | |---|---| | `model-0000*-of-00004.safetensors` + configs | bf16, transformers (`Qwen3_5ForConditionalGeneration`); B4-9B's MTP head is included unchanged | | `RaifuWars-Riko-B4-9B-Q8_0.gguf` | llama.cpp, Q8_0, without the MTP head; fits a 16 GB card with room for context | ```bash llama-server -m RaifuWars-Riko-B4-9B-Q8_0.gguf -c 24576 -ngl 99 -fa on --jinja ``` Use it with thinking off, as it was trained and measured. ## Does merging cost anything? For an earlier B4-9B adapter, merging and then quantizing to Q8_0 lost most of the gain, so this one was measured both ways. 838 held-out narrator prompts, two lines each (1,676 lines), seed 99; "good" means the line passed the rule checker and the judge. | narrator | good | hard fail | judge flag | judge score | |---|---|---|---|---| | B4-9B Q8_0 | 57.2% | 16.0% | 22.6% | 3.63 | | B4-9B Q8_0 + the f16 LoRA at runtime | 92.1% | 0.9% | 3.9% | 3.98 | | **this model, Q8_0 GGUF** | **91.3%** | **0.5%** | **4.2%** | **3.98** | The 0.8-point gap between the last two is within noise at 1,676 lines. On stream she currently runs the second row (B4-9B Q8_0 with the LoRA at runtime); this model is the same narrator as a single file.