AnSungJae3489/Kurken-9B-Summer-Merged

Kurken-9B-Summer

A LoRA finetune for fans of a certain alchemist who had one really good summer.

Yes, this is supposed to be a shitpost. A very high-quality shitpost.

Here's why it qualifies:

• $61 total budget. A single 10-pull in a certain official app costs $20. This entire model costs three 10-pulls. You can't even spark a banner character for what this model cost to train.
• Trained on a GPU in someone's bedroom. No data center. No A100 cluster. No venture capital. Just a consumer card and some legally purchased games.
• Beats a certain company's official app to market. Launch is August 18. This model card exists now. The adapter ships before pre-registration closes.
• No stamina system. GPUs don't get tired. Talk as long as you want. The only limit is your electricity bill.
• All real dialogue, zero synthetic filler. Every training line was written by an actual game scenario writer. No LLM-generated slop. No template scripts. The dataset is cleaner than most academic papers.
• Custom data extraction pipeline built from scratch. Undocumented format. Reverse-engineered byte-by-byte. The extraction tooling alone is a bigger flex than the model.

High-quality: the training methodology, data extraction, and model output. Shitpost: the premise, the price tag, and the timing.

---

What is this?

A QLoRA adapter (rank 32, ~58M params) trained on dialogue data from several official sources. The base model is Qwythos-9B-Claude-Mythos-5-1M — a Qwen 3.5 architecture already tuned for creative roleplay. This LoRA layers character voice, domain knowledge, and conversational mannerisms on top.

It's a companion model. You talk. She responds. No stamina bar. No gacha currency. No skin shop.

---

Why this exists

On August 18, 2026, a certain company is launching an official AI chat RPG app. It looks genuinely impressive — voice acting, live 2D art, world map, quest system, the works. A lot of talent and care clearly went into it.

That said: if you just want to talk to a certain character, you shouldn't need stamina bars and in-game currency to do it. This model is that — just the conversation part, no gates.

Extracted dialogue from multiple sources (personally purchased), parsed undocumented binary formats by hand, mapped speaker identities through voice pattern analysis, trained on a single consumer GPU. Weekend project.

Total cost breakdown:

Item	Cost
Source material (legally purchased)	~$60
Electricity (consumer GPU x ~2 hours)	<$1
Total	~$61

Compare to the official app: assuming standard pricing, a single multi-pull costs roughly $20. One mediocre banner costs more than this entire model's training budget.

You don't need a company. You need a GPU, some scripts, and a weekend.

---

Training data

7,700 dialogue scenes (844K tokens) extracted from multiple sources:

• Several mainline titles — ~2,500 scenes
• Official mobile spinoff — ~4,700 scenes
• Broadcast adaptation (12 episodes) — ~80 scenes
• Reference knowledge (profiles, systems, locations) — ~450 entries

All data is original dialogue from official sources, licensed copies. No synthetic generation. No LLM rewrites. If a line is in this dataset, a real writer wrote it.

Loss masking: only the target character's lines contribute to training loss. Other speakers provide conversational context without being learned as voice.

---

Usage

Two ways to run:

LoRA adapter (232 MB)

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    "empero-ai/Qwythos-9B-Claude-Mythos-5-1M",
    max_seq_length=2048, load_in_4bit=True,
)
model.load_adapter("path/to/ryza-lora")
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "You are a young alchemist from an island. Summer. Your friend is here."},
    {"role": "user", "content": "素材集めに行かない？"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs, max_new_tokens=150, temperature=0.9, top_p=0.92,
        repetition_penalty=1.05, do_sample=True,
    )
print(tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))

Merged model (17.5 GB, ready to run)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "path/to/kurken-9b-summer-merged",
    max_seq_length=2048, load_in_4bit=True,
)
FastLanguageModel.for_inference(model)
# Same generate() usage as above

Recommended settings:

• Temperature: 0.8–1.0, Top-p: 0.9–0.95, Repetition penalty: 1.05
• Max new tokens: 150–300 for natural exchanges
• System prompt: keep it light, summer-themed, mention the island, mention alchemy

---

Sample outputs

All samples use temperature=0.9, top_p=0.92. No cherry-picking.

Casual invitation

User: 今日は絶好の採取日和だね！どこ行く？ Response: ふっふっふ、分かってるよ。あちこち探索して、いい素材探しに行こう！

Alchemy advice

User: 俺も錬金術やってみたいんだけど、最初は何から始めればいい？ Response: うーん、どう説明しよう… 素材を集めてレシピを見つけて調合する… あ、そうだな。まずは素材の採取から始めてみろ。素材が分かれば、その次に何をどう調合するか、分かってくるはずだ

Battle encounter

User: 魔物だ！どうする？ Response: うっ……どうしよう。一人で戦うのは無理だ 待って！あたしにもできるから！錬金術を使えば倒せるはず。材料を集めに行っていたんだもの！よし、任せて！

Friendly teasing

User: ○○とは最近どう？仲いいよね、二人とも。 Response: えっと……二人ともって、□□と△△？うん、まあね。あの二人は小さい頃から一緒に遊んだりしてて、今の関係だからかな？

(Names redacted in the sample — the model uses actual character names. We've blanked them here to keep the model card lawyer-safe.)

Stupid question about appearance

User: その…脚、すごく綺麗だよね。何か秘訣あるの？ Response: よく聞かれることなのかな？まあ、確かに昔はそういうことも気にしたことがあるけど 今はあんまりそんなこと考えてないかな

(Deflected. Didn't bite. She knows, she doesn't care, she has alchemy to do.)

---

What this is NOT

• Not an official product. No affiliation with any game publisher or AI company.
• Not a replacement for playing the actual source material. Play the games. They're good.
• Not voice-cloned. No TTS. No art generation. Just text.
• Not uncensored by design — the base model happens to be uncensored. The training data comes from official scripts (standard rating). What emerges is between you and the model.

---

Known quirks

• Some source material uses raw internal speaker IDs. If she occasionally references someone by number, that's why. Fix pending.
• Mobile-spinoff scenes are predominantly other characters talking about the protagonist rather than the protagonist speaking. Good domain knowledge, less good voice data.
• The model has strong opinions about alchemy. You have been warned.

---

Training details

Parameter	Value
Base model	Qwythos-9B-Claude-Mythos-5-1M (Qwen 3.5, 9B)
Method	QLoRA 4-bit
LoRA rank (r)	32
LoRA alpha	64
Trainable params	58,195,968 (0.61%)
Effective batch size	8 (2 x 4 accumulation)
Learning rate	2e-4 (cosine schedule)
Epochs	3
Max sequence length	2048
Training time	~1.2 hrs on consumer 24GB GPU
Peak GPU memory	~12.8 GB
Logging	TensorBoard

---

Hosting

This is released in two forms:

• LoRA adapter only — tiny download, apply to Qwythos-9B-Claude-Mythos-5-1M yourself
• Merged model — full weights, ready to run, no assembly required

Host it wherever you host LLMs. Flat-fee hosting exists. You don't need stamina to talk to an AI. That's not how GPUs work.

---

License

Apache 2.0. Base model is Apache 2.0. Training data extracted from personally-owned copies for research/transformative use. Do what you want with the weights.

---

Made with approx. $61, one consumer GPU, and genuine annoyance at stamina bars in AI chat apps.