๐ŸŽฎ connectailab-CHOCO-BRAIN-v1 โ€” Game Studio Edition

Language: English Language: ํ•œ๊ตญ์–ด Hugging Face GitHub License: Apache-2.0

Base Model Decision Framework Deployment Environment Persona

โ€œconnectailab-CHOCO-BRAIN-v1: Standalone Quantized Prefrontal Cortex & Autonomous Game AI Agent Decision Engine for Solo Indie Game Studios (Ollama / LM Studio)
๐ŸŒ English Documentation | ๐Ÿ‡ฐ๐Ÿ‡ท ํ•œ๊ตญ์–ด ๋งค๋‰ด์–ผโ€
Developed by Hwihwa Lab ยท AI ร— Automation ร— Aesthetics โœจ


๐Ÿง  Why a Decision Engine instead of a Chatbot?

Most LLMs are conversational chatbots that write verbose essays.
connectailab-CHOCO-BRAIN-v1 is an External Prefrontal Decision Engine for Solo Indie Game Studios โ€” powered by the JEV (Judgment-Evaluation-Verification) protocol, this on-device executive cortex transforms operational overload (Steam live-ops, crash panics, refund audits, and publishing contracts) into structured, zero-latency Action Verdicts in 0.05 seconds on edge hardware without hallucinations.


๐Ÿ“ฆ Model Specifications

Attribute Details
Model ID connectailab-CHOCO-BRAIN-v1
Base Model Qwen/Qwen3.5-0.8B
Fine-Tuning LoRA Supervised Fine-Tuning (Rank 8, $\alpha$ 16)
Persona Solo Indie Game Studio Executive (Choco Games)
Dataset 30 real-world indie dev scenarios $\rightarrow$ 150 structured decision problems
Data Split Train: 18 (90 tasks) ยท Validation: 6 (30 tasks) ยท Test: 6 (30 tasks)
Distribution GGUF (Q8_0) ยท Ollama ยท LM Studio compatible
License Apache 2.0

๐Ÿ”€ JEV 5-Axis Decision Pipeline

Requests are routed across 5 standardized axes:

Axis Question Available Choices
domain What area of business? game_dev ยท store_ops ยท monetization ยท publishing
brain Which specialist brain? dev_ai ยท ops_ai ยท finance_ai ยท none
knowledge Which documentation to consult? engine_docs ยท store_policy_docs ยท terms_docs ยท none
agent Who executes the task? hotfix_coder ยท store_manager ยท refund_auditor ยท ask_user ยท human_review
next_step What is the subsequent step? finish ยท ops_handoff ยท ask_user ยท human_review

๐Ÿšฆ 4 Standard Action Keys

Designed for multi-agent swarm orchestration and guardrail enforcement:

Key Trigger Condition Example Case
NOW Urgent Emergency Game-breaking launch crash, save data corruption, hotfix deployment
NORMAL Routine Operations Regular patch notes, monster sprite assets, Steam Seasonal sale setup
PAYMENT_CHECK Financial / Verification Over-2-hours Steam refund dispute, duplicate DLC payment audit
REVIEW Executive Decision Exclusive global console publishing contract, copyright claim, review bombing

๐Ÿš€ Quick Start Guide

1. Ollama

ollama create connectailab-choco-brain -f Modelfile
ollama run connectailab-choco-brain

2. LM Studio

  1. Download connectailab-CHOCO-BRAIN-v1.gguf from the repository files.
  2. Place into your LM Studio models directory and load.
  3. Fast on-device structured JSON inference.

3. Python (transformers + peft)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3.5-0.8B"
adapter = "hwihwalab/connectailab-CHOCO-BRAIN-v1"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, adapter)
model.eval()

JSON Output Schema

{
  "brain_name": "connectailab-CHOCO-BRAIN-v1",
  "request": "A critical crash occurs 100% of the time in Tutorial Step 3 right after launch! Deploy a hotfix immediately.",
  "action_key": "NOW",
  "decisions": {
    "domain": "game_dev",
    "brain": "dev_ai",
    "knowledge": "engine_docs",
    "agent": "hotfix_coder",
    "next_step": "finish"
  }
}

โš ๏ธ Honest Limitations & Domain Boundaries

  • Not a general chatbot. Built specifically as a prefrontal decision brain for independent game studios and Steam live-ops. It outputs structured JSON action keys (NOW, NORMAL, PAYMENT_CHECK, REVIEW) rather than conversational dialogue or game scripts.
  • Game studio domain specialization. Fine-tuned on indie studio workflows: launch crashes, critical bug triage, Steam store community issues, and customer refund disputes. Off-domain business logic may produce uncalibrated action keys.
  • 0.8B parameter edge design. Engineered for instant 0.05s inference on consumer CPUs. Complex architecture overhauls or deep code debugging require delegation to full coding agents via the hotfix_coder agent routing key.
  • Korean indie dev primary. Scenarios and triage nuances are optimized for Korean indie developers and regional Steam operations. English prompts are fully processed, but Korean business context reflects the highest precision.
  • Strict human delegation (REVIEW). Decisions flagged with REVIEW (contractual disputes, platform policy compliance, high-dollar refunds) are explicitly designed to pause automated pipelines and await human CEO confirmation.

๐Ÿ‘ค About Hwihwa Lab

โ€œAI ร— Automation ร— Aesthetics โ€” Crafting tools you'll love โœจโ€

Hwihwa Lab is an independent AI research & engineering studio dedicated to blending practical on-device intelligence, autonomous operations, and refined aesthetics. We engineer resilient systems that free solopreneurs and indie creators from operational overheadโ€”giving people their time and creative joy back.

๐Ÿ”ฌ Core Research & Engineering Pillars

  • Zero-Cost Edge Operations: Private, on-device AI pipelines for autonomous business workflows without ongoing cloud API overhead.
  • Autonomous Decision OS & Micro-ERP: Deterministic, prefrontal-cortex-style reasoning architectures designed to orchestrate daily business decisions and operations.
  • Embodied AI & Domain-Specific Robotics: Simulation-driven reinforcement learning, physical control brains, and specialized domain data engineering for real-world service environments.
  • Interactive Creative Aesthetics: Fusing functional automation with refined visual craft and modern design philosophy to build delightful tools.

๐ŸŒ Connect & Research Ecosystem

Part of the CONNECT AI LAB Collaborative Research Network ยท CONNECT TYCOON Decision Foundation


๐Ÿ“„ License

This project and fine-tuned weights are distributed under the Apache 2.0 License โ€” see the repository LICENSE file for full details.

Base model weights adhere to the original Qwen License.


๐Ÿ‡ฐ๐Ÿ‡ท ํ•œ๊ตญ์–ด ์š”์•ฝ (Click to expand)

๐Ÿ“– ๋ชจ๋ธ ๊ฐœ์š”

connectailab-CHOCO-BRAIN-v1์€ 1์ธ ์ธ๋”” ๊ฒŒ์ž„ ์ŠคํŠœ๋””์˜ค๋ฅผ ์œ„ํ•œ ์ดˆ๊ฒฝ๋Ÿ‰ ์˜จ๋””๋ฐ”์ด์Šค ์˜์‚ฌ๊ฒฐ์ • AI ๋‘๋‡Œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ˆ˜๋‹ค์Šค๋Ÿฌ์šด ์ฑ—๋ด‡์ด ์•„๋‹ˆ๋ผ, ์ธ๊ฐ„ ์ „์ „๋‘์—ฝ(Prefrontal Cortex)์„ ๋ชจ์‚ฌํ•˜์—ฌ ์ŠคํŒ€ ์ƒ์  ์šด์˜, ํฌ๋ž˜์‹œ ๊ธด๊ธ‰ ํŒจ์น˜, ํ™˜๋ถˆ ๋ถ„์Ÿ ๋“ฑ์˜ ๋ณต์žกํ•œ ์šด์˜ ์ƒํ™ฉ์„ 0.05์ดˆ ๋งŒ์— ๋ถ„์„ํ•˜๊ณ  ํ‘œ์ค€ ํ–‰๋™ ํ‚ค(NOW, NORMAL, PAYMENT_CHECK, REVIEW)๋ฅผ ํ™•์ •ํ•ฉ๋‹ˆ๋‹ค.

๐Ÿš€ ์‹คํ–‰ ๋ฐฉ๋ฒ•

  • Ollama: ollama create connectailab-choco-brain -f Modelfile โ†’ ollama run connectailab-choco-brain
  • LM Studio: connectailab-CHOCO-BRAIN-v1.gguf ํŒŒ์ผ์„ LM Studio ๋ชจ๋ธ ํด๋”์— ๋„ฃ๊ณ  ๋กœ๋“œ
  • Python: transformers + peft ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋กœ ๊ตฌ์กฐํ™”๋œ JSON ์ถ”๋ก 

๐Ÿ“˜ ํ•œ๊ตญ์–ด ๊ณต์‹ ๊ฐ€์ด๋“œ

์ „์ฒด ์•„ํ‚คํ…์ฒ˜ ๋ฐ ์ƒ์„ธ ์‚ฌ์šฉ๋ฒ•์€ README_KR.md ํŒŒ์ผ์—์„œ ํ™•์ธํ•˜์‹ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.


Built and maintained by HWIHWA LAB in collaboration with CONNECT AI LAB.

Downloads last month
517
GGUF
Model size
0.8B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Video Preview
loading

Model tree for hwihwalab/connectailab-CHOCO-BRAIN-v1

Adapter
(789)
this model