| --- |
| tags: |
| - ml-intern |
| --- |
| # 🤖 Desktop Agent Autónomo (Sin Censura) |
|
|
| Agente de escritorio autónomo con VLM multimodal sin censura. |
|
|
| ## Arquitectura |
|
|
| ``` |
| 👁️ OJOS → pyautogui.screenshot() → Captura pantalla |
| 🧠 CEREBRO → Qwen3.5-35B-A3B-abliterated → Piensa y decide |
| 🖐️ MANOS → pyautogui → Ejecuta acciones |
| 📚 MEMORIA → DPO online → Aprende de interacciones |
| ``` |
|
|
| ## Modelos Soportados (Sin Censura) |
|
|
| | Modelo | Tamaño | VRAM (4-bit) | Tipo | Link | |
| |--------|--------|--------------|------|------| |
| | **Qwen3.5-35B-A3B-abliterated** ⭐ | 35B/3B activos | ~16GB | MoE | [HF](https://hf.co/huihui-ai/Huihui-Qwen3.5-35B-A3B-abliterated) | |
| | Qwen3.6-27B-abliterated | 27B | ~27GB | Dense | [HF](https://hf.co/wangzhang/Qwen3.6-27B-abliterated) | |
| | Gemma-4-26B-A4B-abliterated | 26B/4B activos | ~14GB | MoE | [HF](https://hf.co/jenerallee78/gemma-4-26B-A4B-it-ara-abliterated) | |
|
|
| ## Instalación |
|
|
| ```bash |
| pip install -r requirements.txt |
| ``` |
|
|
| ## Uso |
|
|
| ### 1. Ejecutar agente |
|
|
| ```bash |
| python agent.py --task "Open Chrome and search for AI news" --steps 20 |
| ``` |
|
|
| ### 2. Entrenar con DPO (aprendizaje) |
|
|
| Primero el agente interactúa y guarda logs. Luego: |
|
|
| ```bash |
| python train_dpo.py --epochs 3 --lr 5e-7 |
| ``` |
|
|
| ### 3. Usar modelo entrenado |
|
|
| ```bash |
| python agent.py --model "Matzan/desktop-agent-dpo" --task "New task" |
| ``` |
|
|
| ## Acciones Soportadas |
|
|
| - `click(x, y)` — Click en coordenadas normalizadas (0-1) |
| - `type("text")` — Escribe texto |
| - `key("enter")` — Presiona tecla |
| - `scroll(x, y, "down")` — Scroll en posición |
| - `done("reason")` — Termina tarea |
| - `fail("reason")` — No puede completar |
|
|
| ## ⚠️ Seguridad |
|
|
| - `pyautogui.FAILSAFE = True` — Mueve mouse a esquina superior izquierda para abortar |
| - El agente puede interactuar con tu desktop real. Úsalo con precaución. |
|
|
| ## Pipeline de Aprendizaje |
|
|
| ``` |
| 1. Agente interactúa → Guarda (screenshot, acción, reward) |
| 2. DPO: compara acciones exitosas vs fallidas |
| 3. Reentrena modelo |
| 4. Repite |
| ``` |
|
|
| <!-- ml-intern-provenance --> |
| ## Generated by ML Intern |
|
|
| This model repository was generated by [ML Intern](https://github.com/huggingface/ml-intern), an agent for machine learning research and development on the Hugging Face Hub. |
|
|
| - Try ML Intern: https://smolagents-ml-intern.hf.space |
| - Source code: https://github.com/huggingface/ml-intern |
|
|
| ## Usage |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| model_id = 'Matzan/desktop-agent-uncensored' |
| tokenizer = AutoTokenizer.from_pretrained(model_id) |
| model = AutoModelForCausalLM.from_pretrained(model_id) |
| ``` |
|
|
| For non-causal architectures, replace `AutoModelForCausalLM` with the appropriate `AutoModel` class. |
|
|
