{ "cells": [ { "cell_type": "markdown", "id": "18b82c6b-10dc-4d94-b8dc-592ff011ce2b", "metadata": {}, "source": [ "# Meeting minutes creator\n", "\n", "https://colab.research.google.com/drive/13wR4Blz3Ot_x0GOpflmvvFffm5XU3Kct?usp=sharing\n", "\n", "## **Week 3 task.**\n", "Create your own tool that generates synthetic data/test data. Input the type of dataset or products or job postings, etc. and let the tool dream up various data samples.\n", "\n" ] }, { "cell_type": "code", "execution_count": null, "id": "e9289ba7-200c-43a9-b67a-c5ce826c9537", "metadata": {}, "outputs": [], "source": [ "# imports\n", "import gradio as gr, requests, json, time, os, torch\n", "from transformers import pipeline, set_seed\n", "from functools import partial\n", "from openai import OpenAI, APIError, AuthenticationError\n", "from google.colab import drive, userdata\n", "from huggingface_hub import login\n", "from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig\n", "\n", "# Sample user_prompt = \"a list of student profiles with full name, email, course studied, and GPA for each of 6 semesters, and a CGPA for the 6 semesters\"\n", "\n", "# Sign in to HuggingFace Hub\n", "hf_token = userdata.get('HF_TOKEN')\n", "login(hf_token, add_to_git_credential=True)\n", "\n", "# Sign in to OpenAI using Secrets in Colab\n", "openai_api_key = userdata.get('OPENAI_API_KEY')\n", "\n", "# Initialize client\n", "try:\n", " openai = OpenAI(api_key=openai_api_key)\n", "except Exception as e:\n", " openai = None\n", " print(f\"OpenAI client not initialized: {e}\")\n", "\n", "# Constants\n", "GPT_MODEL = \"gpt-3.5-turbo\"\n", "\n", "# Local Llama Model Setup\n", "# Loads a Llama model from Hugging Face for local inference.\n", "# Note: This requires a powerful GPU and specific library installations (e.g., bitsandbytes, accelerate).\n", "LLAMA_MODEL = \"meta-llama/Meta-Llama-3.1-8B-Instruct\"\n", "\n", "try:\n", " # Set up quantization config for efficient memory usage.\n", " # This loads the model in 4-bit precision, significantly reducing VRAM requirements.\n", " quant_config = BitsAndBytesConfig(\n", " load_in_4bit=True,\n", " bnb_4bit_use_double_quant=True,\n", " bnb_4bit_compute_dtype=torch.bfloat16,\n", " bnb_4bit_quant_type=\"nf4\"\n", " )\n", "\n", " # Load the tokenizer and model.\n", " tokenizer = AutoTokenizer.from_pretrained(LLAMA_MODEL)\n", " model = AutoModelForCausalLM.from_pretrained(\n", " LLAMA_MODEL, \n", " device_map=\"auto\", \n", " quantization_config=quant_config,\n", " trust_remote_code=True\n", " )\n", " \n", " # Set the model to evaluation mode for inference.\n", " model.eval()\n", "\n", "except Exception as e:\n", " model = None\n", " tokenizer = None\n", " print(f\"Failed to load local Llama model: {e}\")\n", "\n", "\n", "def generate_with_llama(user_prompt: str, num_samples: int = 5):\n", " \"\"\"\n", " Generates synthetic data using a local Llama model.\n", " Return a JSON string.\n", " \"\"\"\n", " if not model or not tokenizer:\n", " return json.dumps({\"error\": \"Llama model not loaded. Check model paths and hardware compatibility.\"}, indent=2)\n", "\n", " # Llama 3.1 uses a specific chat template for conversation formatting.\n", " messages = [\n", " {\"role\": \"system\", \"content\": f\"You are a data generation assistant. Generate a JSON array of exactly {num_samples} objects based on the user's request. The output must be valid JSON only, without any other text or formatting.\"},\n", " {\"role\": \"user\", \"content\": user_prompt}\n", " ]\n", "\n", " try:\n", " inputs = tokenizer.apply_chat_template(messages, return_tensors=\"pt\").to(\"cuda\")\n", "\n", " outputs = model.generate(inputs, max_new_tokens=2000, do_sample=True, top_p=0.9, temperature=0.7)\n", "\n", " # Decode the generated tokens.\n", " response_text = tokenizer.decode(outputs[0], skip_special_tokens=True)\n", "\n", " # Extract only the assistant's part from the complete chat history.\n", " assistant_start = \"<|eot_id|><|start_header_id|>assistant<|end_header_id|>\\n\\n\"\n", " if assistant_start in response_text:\n", " response_text = response_text.split(assistant_start)[-1]\n", " \n", " # Parse the JSON and return it.\n", " parsed_json = json.loads(response_text)\n", " return json.dumps(parsed_json, indent=2)\n", "\n", " except Exception as e:\n", " return json.dumps({\"error\": f\"An error occurred during local model generation: {e}\"}, indent=2)\n", "\n", "\n", "\n", "def generate_with_gpt(user_prompt: str, num_samples: int = 5):\n", " \"\"\"\n", " Generates synthetic data using OpenAI's GPT.\n", " Return a JSON string.\n", " \"\"\"\n", " if not openai:\n", " return json.dumps({\"error\": \"OpenAI client not initialized. Please check your API key.\"}, indent=2)\n", "\n", " try:\n", " response = openai.chat.completions.create(\n", " model=GPT_MODEL,\n", " messages=[\n", " {\"role\": \"system\", \"content\": f\"You are a data generation assistant. Generate a JSON array of exactly {num_samples} objects based on the user's request. The output must be valid JSON only, without any other text or formatting.\"},\n", " {\"role\": \"user\", \"content\": user_prompt}\n", " ],\n", " response_format={\"type\": \"json_object\"}\n", " )\n", " \n", " json_text = response.choices[0].message.content\n", " return json_text\n", " except APIError as e:\n", " return json.dumps({\"error\": f\"Error from OpenAI API: {e.body}\"}, indent=2)\n", " except Exception as e:\n", " return json.dumps({\"error\": f\"An unexpected error occurred: {e}\"}, indent=2)\n", "\n", "\n", "def generate_data(user_prompt, model_choice):\n", " \"\"\"\n", " Wrapper function that calls the appropriate generation function based on model choice.\n", " \"\"\"\n", " if not user_prompt:\n", " return json.dumps({\"error\": \"Please provide a description for the data.\"}, indent=2)\n", "\n", " if model_choice == f\"Hugging Face ({LLAMA_MODEL})\":\n", " return generate_with_llama(user_prompt)\n", " elif model_choice == f\"OpenAI ({GPT_MODEL})\":\n", " return generate_with_gpt(user_prompt)\n", " else:\n", " return json.dumps({\"error\": \"Invalid model choice.\"}, indent=2)\n", "\n", "# Gradio UI\n", "with gr.Blocks(theme=gr.themes.Soft(), title=\"Synthetic Data Generator\") as ui:\n", " gr.Markdown(\"# Synthetic Data Generator\")\n", " gr.Markdown(\"Describe the type of data you need, select a model, and click 'Generate'.\")\n", "\n", " with gr.Row():\n", " with gr.Column(scale=3):\n", " data_prompt = gr.Textbox(\n", " lines=5,\n", " label=\"Data Prompt\",\n", " placeholder=\"e.g., a list of customer profiles with name, email, and a favorite product\"\n", " )\n", " \n", " with gr.Column(scale=1):\n", " model_choice = gr.Radio(\n", " [f\"Hugging Face ({LLAMA_MODEL})\", f\"OpenAI ({GPT_MODEL})\"],\n", " label=\"Choose a Model\",\n", " value=f\"Hugging Face ({LLAMA_MODEL})\"\n", " )\n", " \n", " generate_btn = gr.Button(\"Generate Data\")\n", " \n", " with gr.Row():\n", " output_json = gr.JSON(label=\"Generated Data\")\n", " \n", " # Click trigger\n", " generate_btn.click(\n", " fn=generate_data,\n", " inputs=[data_prompt, model_choice],\n", " outputs=output_json\n", " )\n", "\n", "ui.launch(inbrowser=True, debug=True)\n" ] }, { "cell_type": "code", "execution_count": null, "id": "cd2020d3", "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.11.7" } }, "nbformat": 4, "nbformat_minor": 5 }