Modelpublic

onevloth/pc-agent-dispatcher-gemma4-e2b

sourceHugging Facegemmaupdated 3mo agoView on Hugging Face

0likes10downloads

Model Card

PC-Agent Dispatcher — Gemma 4 E2B (QLoRA LoRA adapter)

A LoRA adapter that turns Gemma 4 E2B into the dispatcher (router) for **PC-Agent**: it reads a natural-language request and emits one JSON tool-call for controlling a Windows PC.

json

{"action": "tool", "tool": "open_app", "args": {"app": "notepad"}}
{"action": "plan", "steps": [{"tool": "open_app", "args": {"app": "notepad"}},
                             {"tool": "type_text", "args": {"text": "hello"}}]}
{"action": "chat"}

🔗 Full project, desktop app, and demos on GitHub: [github.com/AlfatihRabbani/pc-agent](https://github.com/AlfatihRabbani/pc-agent)

Install & run the full app

bat

git clone https://github.com/AlfatihRabbani/pc-agent
cd pc-agent
scripts\setup.bat                                              :: venv + CUDA torch + deps
python scripts\download_models.py                             :: base E2B (~10 GB)
.venv\Scripts\hf download onevloth/pc-agent-dispatcher-gemma4-e2b --local-dir models\dispatcher-final
PC-Agent.vbs                                                  :: launch the desktop app

Full instructions (incl. the optional 12B chat model via Ollama) are in the repo README.

Details

—Base: huihui-ai/Huihui-gemma-4-E2B-it-abliterated (load in 4-bit / NF4).
—Method: QLoRA, r=16, α=32, attn+MLP projections; 2 epochs (~3,196 steps), seq 512, batch 1 × accum 16, on a single RTX 3080 Ti (12 GB).
—Data: ~25.8k examples — function-calling (xLAM/Hermes-style) + ~2k synthetic Windows-action examples generated from the agent's own tool registry.
—Use it for routing; chat with the base model (adapter disabled).

Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

base = "huihui-ai/Huihui-gemma-4-E2B-it-abliterated"
q = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                       bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, quantization_config=q, device_map="auto")
model = PeftModel.from_pretrained(model, "onevloth/pc-agent-dispatcher-gemma4-e2b")

License

Derivative of Google Gemma 4 — governed by the Gemma Terms of Use. The PC-Agent code is MIT.