Files
Table with columns: file, use| file | use |
|---|
adapter_model.safetensors, adapter_config.json | PEFT adapter (r 32, alpha 64, attention and MLP projections) |
RaifuWars-Riko-B4-9B-LoRA-f16.gguf | the same adapter for llama.cpp: llama-server -m B4-9B-Q8_0.gguf --lora RaifuWars-Riko-B4-9B-LoRA-f16.gguf |
All numbers below were measured the second way: B4-9B Q8_0 with this f16 LoRA loaded at
runtime, thinking off.
What it was trained on
Her prompts are the live game: whose turn it is, what just happened, the standings, her own
options, and a short read of the board (how far she is from the next tier, her best shot, who is
threatening her). Prompts come from real Raifu Wars matches played by her policy.
teacher B5-27B + Riko's narrator LoRA, one line for each of 6,049 narrator promptsfilter a rule checker (format, length, pronouns, invented facts, stock phrases, echoing the report back) and a 27B judge (faithful to the game, owns her moves, score >= 4/5); the same line kept at most 3 timesstudent B4-9B, SFT on the 5,120 lines that passed (85%), loss on her line only 2 epochs, lr 1e-4 cosine, LoRA r 32 / alpha 64, max length 3072
The training lines are the teacher's own words, unedited: the student learns the teacher's best
lines and none of its failures. The prompts and lines are not released.
Numbers
838 held-out narrator prompts, two lines each (1,676 lines), seed 99. "Good" means the line passed
the rule checker and the judge.
Table with columns: narrator, good, hard fail, judge flag, judge score| narrator | good | hard fail | judge flag | judge score |
|---|
| B4-9B | 57.2% | 16.0% | 22.6% | 3.63 |
| B4-9B + an earlier LoRA | 70.7% | 4.5% | 19.6% | 3.69 |
| B5-27B + narrator LoRA (the teacher) | 89.7% | 1.7% | 6.0% | 3.97 |
| this LoRA | 92.1% | 0.9% | 3.9% | 3.98 |
What it fixed in the base 9B, counted over the same 1,676 lines:
Table with columns: problem, B4-9B, this LoRA| problem | B4-9B | this LoRA |
|---|
| wrong pronouns (every character in the game is a girl) | 88 | 0 |
| echoing the game report back | 56 | 0 |
| stock phrases | 43 | 1 |
| number dumps | 27 | 1 |
| invented hits | 21 | 6 |
| over-long lines | 80 | 3 |
Limitations
- The judge is the same kind of model that filtered the training lines. The prompts are new
to the model, but part of the gain is learning what this judge likes. It has also been watched
in a live narrated match and held up, but that is one match.
- Pronoun nouns slip through. In its first live match it called another player "a proper
gentleman". The checker missed nouns like that until then; it catches them now.
- It is built for one character and one prompt format. Outside Riko's narrator prompts it is
B4-9B with a habit of short, in-character game commentary.