Upload demo_colab.ipynb with huggingface_hub

demo_colab.ipynb

@@ -0,0 +1,118 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Tiny Browser Planner — Live Demo\n",
+    "**Reason-First model** | MiniCPM5-1B + LoRA\n",
+    "\n",
+    "Run all cells below. At the end, a public URL will appear — click it to use the demo."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Install dependencies\n",
+    "!pip install unsloth gradio datasets transformers torch --quiet\n",
+    "!pip install bitsandbytes accelerate --quiet"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import re, torch\n",
+    "from unsloth import FastLanguageModel\n",
+    "import gradio as gr\n",
+    "\n",
+    "MODEL_ID = \"Georgefifth/tiny-browser-planner-reason\"\n",
+    "\n",
+    "print(\"Loading model...\")\n",
+    "model, tokenizer = FastLanguageModel.from_pretrained(\n",
+    "    MODEL_ID, max_seq_length=2048, load_in_4bit=True, dtype=torch.bfloat16,\n",
+    ")\n",
+    "model = FastLanguageModel.get_peft_model(model, r=16,\n",
+    "    target_modules=['q_proj','k_proj','v_proj','o_proj','gate_proj','up_proj','down_proj'],\n",
+    "    lora_alpha=16)\n",
+    "model.load_adapter(MODEL_ID, 'default')\n",
+    "FastLanguageModel.for_inference(model)\n",
+    "print(\"Loaded!\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def predict(task, history_text):\n",
+    "    history = [l.strip() for l in history_text.strip().split(chr(10)) if l.strip()]\n",
+    "    hist_str = chr(10).join(history)\n",
+    "    msgs = [\n",
+    "        {'role': 'system', 'content': 'You are a browser planner. First reason about the situation, then output the next action.'},\n",
+    "        {'role': 'user', 'content': f'Task: {task}\\n\\nHistory:\\n{hist_str}\\n\\nWhat is the next action?'},\n",
+    "    ]\n",
+    "    prompt = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)\n",
+    "    inputs = tokenizer(prompt, return_tensors='pt').to('cuda')\n",
+    "    input_len = inputs['input_ids'].shape[1]\n",
+    "    outs = model.generate(**inputs, max_new_tokens=64, temperature=0.01, do_sample=False, pad_token_id=tokenizer.eos_token_id)\n",
+    "    output = tokenizer.decode(outs[0][input_len:], skip_special_tokens=True).strip()\n",
+    "    reason_m = re.search(r'Reason:\\s*(.+?)(?:\\n|\\$)', output)\n",
+    "    action_m = re.search(r'Action:\\s*(\\S+)', output)\n",
+    "    return (reason_m.group(1).strip() if reason_m else '?'), (action_m.group(1).strip().lower() if action_m else '?')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "PRESETS = [\n",
+    "    ('Find Apple stock price', '[search] Search completed.\\n[open_page] Price displayed prominently at $198'),\n",
+    "    ('Find Apple stock price', '[search] Search completed.\\n[open_page] Product review page, not stock data'),\n",
+    "    ('Find CEO of OpenAI', '[search] Search completed.\\n[open_page] API pricing page, not CEO info'),\n",
+    "    ('Find AWS EC2 pricing', '[search] Search completed.\\n[open_page] Pricing behind login wall'),\n",
+    "    ('Find Python 3.12 release date', '[search] Search completed.\\n[open_page] Release date listed on official page'),\n",
+    "    ('Find Tesla Model Y price', '[search] Search completed.\\n[open_page] Shows Model 3 pricing, not Model Y'),\n",
+    "]\n",
+    "\n",
+    "with gr.Blocks(title='Tiny Browser Planner', theme='soft') as demo:\n",
+    "    gr.Markdown('''# Tiny Browser Planner\n",
+    "**Reason-First** — MiniCPM5-1B + LoRA  |  Actions: search, open_page, extract, refine_search, back, finish''')\n",
+    "    task = gr.Textbox(label='Task', placeholder='Find Apple stock price')\n",
+    "    history = gr.Textbox(label='History (one action per line)', lines=4,\n",
+    "                         placeholder='[search] Search completed.\\n[open_page] Price displayed')\n",
+    "    btn = gr.Button('Predict', variant='primary')\n",
+    "    with gr.Row():\n",
+    "        reason = gr.Textbox(label='Reason', interactive=False, lines=2)\n",
+    "        action = gr.Textbox(label='Action', interactive=False, lines=1)\n",
+    "    btn.click(fn=predict, inputs=[task, history], outputs=[reason, action])\n",
+    "    \n",
+    "    gr.Markdown('### Quick Examples')\n",
+    "    for t, h in PRESETS:\n",
+    "        gr.Button(t, size='sm').click(\n",
+    "            fn=lambda t=t, h=h: (t, h), outputs=[task, history]\n",
+    "        ).then(fn=predict, inputs=[task, history], outputs=[reason, action])\n",
+    "\n",
+    "print('\\n=== Click the URL below to open the demo ===\\n')\n",
+    "demo.launch(share=True, server_name='0.0.0.0')"
+   ]
+  }
+ ],
+ "metadata": {
+  "accelerator": "GPU",
+  "language_info": {
+   "name": "python",
+   "version": "3.10"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}