🟣 JujutsuKaiserver

The Cursed Intelligence. The Canon Oracle.

"Throughout Heaven and Earth, I alone am the honored one."
— Satoru Gojo | and also this model, kind of.

JujutsuKaiserver is a Retrieval-Augmented Generation (RAG) model built for one purpose:
to answer anything and everything about the Jujutsu Kaisen universe — with canon-backed accuracy, zero hallucination tolerance, and the confidence of Unlimited Void.

---

⚡ What It Does

Ask it anything. Techniques. Domains. Arcs. Hidden lore. Character relationships. Cursed Energy mechanics. It retrieves the most relevant passages from a 200+ page wiki corpus, feeds them into a fine-tuned Qwen2.5-1.5B-Instruct backbone, and gives you a clean, grounded answer — not a guess.

Ask This	Get This
"What is Sukuna's Shrine?"	Full technique breakdown with canon context
"How does Mahito's Idle Transfiguration work?"	Soul-level mechanics explained accurately
"What happened in the Shibuya Incident?"	Arc summary backed by wiki chunks
"Who is the strongest Grade 1 sorcerer?"	Ranked answer with sourced reasoning

---

🧠 Architecture

User Query
    │
    ▼
sentence-transformers (all-MiniLM-L6-v2)
    │  [embed query]
    ▼
FAISS Index (jjk_index.faiss)
    │  [top-5 relevant wiki chunks]
    ▼
Qwen2.5-1.5B-Instruct (4-bit)
    │  [context + question → chat template]
    ▼
Canon-grounded Answer

Model Composition

Component	Details
🤖 Base LLM	Qwen/Qwen2.5-1.5B-Instruct (4-bit quantized)
🔢 Embeddings	sentence-transformers/all-MiniLM-L6-v2
📦 Vector Store	FAISS — jjk_index.faiss
📖 Knowledge Base	120+ cleaned JJK Fandom Wiki articles (chunks.txt)
🔧 Pipeline	Custom JujutsuKaiserver class with Qwen chat template

---

🚀 Quick Start

from huggingface_hub import snapshot_download

model_dir = snapshot_download("TinyModels/JujutsuKaiserver")

import sys
sys.path.insert(0, model_dir)
from pipeline import JujutsuKaiserver

bot = JujutsuKaiserver(model_dir=model_dir)

# Ask anything
print(bot.ask("What is Gojo's Domain Expansion called?"))
# → "Infinite Void (無量空処). It..."

⚠️ Requirements: bitsandbytes, GPU with ≥6 GB VRAM. CPU inference works but is slow.

Install Dependencies

pip install transformers bitsandbytes faiss-cpu sentence-transformers huggingface_hub

---

🖥️ Gradio Demo (Optional)

Spin up a local chat UI in seconds:

import gradio as gr
from pipeline import JujutsuKaiserver

bot = JujutsuKaiserver(model_dir="<path_to_downloaded_model>")

def chat(message, history):
    return bot.ask(message)

gr.ChatInterface(
    fn=chat,
    title="🟣 JujutsuKaiserver",
    description="Ask anything about the JJK universe."
).launch()

---

✨ Features

• 🔍 Factual Q&A — Every answer is grounded in retrieved wiki content, not imagination
• 🚫 Hallucination Guard — Model is prompted to say "I don't know" when context is insufficient
• 📚 Deep Coverage — 200+ wiki pages: characters, techniques, domains, arcs, lore
• ⚡ T4-Friendly — 4-bit quantization means it runs on free Colab tiers
• 🤖 Gradio Ready — One-script local demo included out of the box

---

⚠️ Known Limitations

• Recent chapters beyond the scraping date may not be indexed yet
• Ambiguous context can still occasionally produce imperfect answers — being addressed via a feedback loop
• Roleplay mode is possible with a custom system prompt, but this version is optimized for factual retrieval

---

🔮 Roadmap

• Live Feedback Flagging — 👍/👎 votes from the Gradio Space feed a correction dataset automatically
• Self-Correcting Pipeline — Weekly DPO fine-tuning on flagged examples + FAISS index refresh
• Expanded KB — Episode transcripts, manga panels text, community lore
• Streaming Support — Token-by-token output for snappier UX

---

📂 Repo Structure

JujutsuKaiserver/
├── pipeline.py          # Core RAG pipeline class
├── jjk_index.faiss      # FAISS vector index
├── chunks.txt           # Raw wiki knowledge base
├── generation_config.json
└── README.md

---

Built with 🩸 and cursed energy for the JJK community.

Got a question the bot fumbled? Open a Discussion and help us fix it.

TinyModels • QuantaSparkLabs • Apache 2.0