Modelpublic

rakesh0x/qwen3-4b-agentnet-text-lora

sourceHugging Facemitupdated 1mo agoView on Hugging Face

0likes

Model Card

Qwen3-4B AgentNet text LoRA (MLX)

LoRA adapter for `mlx-community/Qwen3-4B-4bit`, trained on text-only Ubuntu steps from xlangai/AgentNet (agentnet_ubuntu_5k.jsonl). No screenshots were used.

This adapter teaches OpenCUA-style Thought / Action / Code (pyautogui) from a written screen observation. It does not add visual grounding.

Training

—Method: MLX LoRA (mlx_lm.lora), rank 8, 8 layers, mask-prompt
—Steps: 400 · batch 1 · lr 1e-5 · max seq 2048
—Data: 2000 AgentNet text steps
—Trainable params: 3.67M (0.091%)
—Val loss: 1.955 → 0.989
—Hardware: Apple M4, peak ~3.9 GB

Use

bash

python3 -m mlx_lm generate \
  --model mlx-community/Qwen3-4B-4bit \
  --adapter-path rakesh0x/qwen3-4b-agentnet-text-lora \
  --prompt "You are a GUI agent. Open LibreOffice Writer. Output Thought, Action, and Code."

python

from mlx_lm import load, generate

model, tokenizer = load(
    "mlx-community/Qwen3-4B-4bit",
    adapter_path="rakesh0x/qwen3-4b-agentnet-text-lora",
)

Final weights: adapters.safetensors. Checkpoints 0000100 … 0000400 are intermediate saves.

If you use AgentNet / OpenCUA, please cite Wang et al., 2025.