Spaces:

---

umyunsang

/

govon-runtime

like 1

Paused

Project: GovOn — Korean Government Civil Complaint AI Assistant

We are a university research team building an AI-powered system to help
Korean government agencies respond to civil complaints more efficiently
and accurately.

Why we need GPU:

• Our base model (EXAONE 4.0-32B-AWQ, ~20GB VRAM) requires at least
  24GB GPU for inference
• We serve Multi-LoRA adapters (civil response + legal citation) via
  vLLM per-request switching
• This is an academic project — we have no commercial budget for GPU

What we built:

• LangGraph approval-gated agent runtime with human-in-the-loop
• 74K civil response training data from AI Hub public datasets
• 270K legal citation training data (precedents + statutes)
• QLoRA fine-tuned adapters on EXAONE 4.0-32B
• CLI tool with streaming status display

Hardware requested: A100 (80GB) or L40S (48GB)

• 32B-AWQ model (~20GB) + Multi-LoRA + KV cache needs 40-48GB
• Used for academic demo and user acceptance testing

GitHub: https://github.com/GovOn-Org/GovOn
Datasets: huggingface.co/datasets/umyunsang/govon-civil-response-data
License: MIT (code), EXAONE 1.2-NC (model)