File size: 1,967 Bytes
c14d436
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
---
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.