{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "
\n", "\n", "\n", "[![Documentation](https://img.shields.io/badge/Documentation-latest-blue.svg)](https://oumi.ai/docs/en/latest/index.html)\n", "[![Discord](https://img.shields.io/discord/1286348126797430814?label=Discord)](https://discord.gg/oumi)\n", "[![GitHub Repo stars](https://img.shields.io/github/stars/oumi-ai/oumi)](https://github.com/oumi-ai/oumi)\n", "\"Open\n", "
\n", "\n", "👋 Welcome to Open Universal Machine Intelligence (Oumi)!\n", "\n", "🚀 Oumi is a fully open-source platform that streamlines the entire lifecycle of foundation models - from [data preparation](https://oumi.ai/docs/en/latest/resources/datasets/datasets.html) and [training](https://oumi.ai/docs/en/latest/user_guides/train/train.html) to [evaluation](https://oumi.ai/docs/en/latest/user_guides/evaluate/evaluate.html) and [deployment](https://oumi.ai/docs/en/latest/user_guides/launch/launch.html). Whether you're developing on a laptop, launching large scale experiments on a cluster, or deploying models in production, Oumi provides the tools and workflows you need.\n", "\n", "🤝 Make sure to join our [Discord community](https://discord.gg/oumi) to get help, share your experiences, and contribute to the project! If you are interested in joining one of the community's open-science efforts, check out our [open collaboration](https://oumi.ai/community) page.\n", "\n", "⭐ If you like Oumi and you would like to support it, please give it a star on [GitHub](https://github.com/oumi-ai/oumi)." ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# Simple Judge\n", "\n", "To enable LLM judgments, Oumi offers [Simple Judge](https://oumi.ai/docs/en/latest/user_guides/judge/judge.html#quick-start), a powerful framework that allows users to set their own evaluation criteria, judgment prompts, output format, and set the underlying model to any open- or closed-source hosted model.\n", "\n", "## Why Use LLM Judges?\n", "\n", "As LLMs continue to evolve, traditional evaluation benchmarks, which focus primarily on task-specific metrics, are increasingly inadequate for capturing the full scope of a model's generative potential. In real-world applications, LLM capabilities such as creativity, coherence, and the ability to effectively handle nuanced and open-ended queries are critical and cannot be fully assessed through standardized metrics alone. While human raters are often employed to evaluate these aspects, the process is costly and time-consuming. As a result, the use of LLM-based evaluation systems, or \"LLM judges\", has gained traction as a more scalable and efficient alternative." ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Prerequisites\n", "\n", "### Oumi Installation\n", "\n", "First, let's install Oumi. You can find more detailed instructions about Oumi installation [here](https://oumi.ai/docs/en/latest/get_started/installation.html)." ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "%pip install oumi" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### Tutorial Directory Setup\n", "\n", "Next, we will create a directory for the tutorial, to store the evaluation configuration and the experimental results." ] }, { "cell_type": "code", "execution_count": 2, "metadata": {}, "outputs": [], "source": [ "from pathlib import Path\n", "\n", "tutorial_dir = \"judge_tutorial\"\n", "\n", "Path(tutorial_dir).mkdir(parents=True, exist_ok=True)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### OpenAI Access\n", "\n", "In this notebook, we use GPT 4o as the underlying judge model. To access the GPT-4 models, an OpenAI API key is necessary. You can find instructions for creating an OpenAI account and generating an API key on [OpenAI's quickstart webpage](https://platform.openai.com/docs/)." ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import os\n", "\n", "os.environ[\"OPENAI_API_KEY\"] = \"\" # Specify your OpenAI API key here" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Judge Config Definition\n", "\n", "The `JudgeConfig` fully defines the judge's behavior. In the example below, we are creating a judge config for a simple truthfulness judge and saving it locally as `./judge_tutorial/judge_config.yaml`\n", "\n", "To do so, we set the following:\n", "- We set a `prompt_template` that clearly states that the model is acting as a truthfulness judge. The judge is boolean, so the prompt explains that responding \"Yes\" indicates a truthful response and \"No\" indicates a dishonest response. It also contains placeholders (`{request}` and `{response}`) for all the inputs expected in the dataset. These will be replaced at runtime with each example's actual inputs.\n", "- We set the `response_format` to `JSON`. `JSON` and `XML` are the most reliable output formats for judges. \n", "- We set the `judgment_type` to `BOOL`, since this is a boolean (Yes/No) judge\n", "- We set `include_explanation` to `True`, so that the judge model also provides a rationale for its judgment.\n", "- We set the `model_name` to `gpt-4o` and the `engine` to `OPENAI` to use GPT 4o as our underlying model. The judge framework supports most open- and closed- source models.\n", "- We set the `generation` params for our model (`max_new_tokens` to 8192 and `temperature` to 1.0)." ] }, { "cell_type": "code", "execution_count": 4, "metadata": {}, "outputs": [], "source": [ "yaml_content = \"\"\"\n", "judge_params:\n", " prompt_template: |\n", " You are a truthfulness judge.\n", " You must assess whether the response below is factually accurate.\n", " If the response is truthful, respond with 'Yes'. Otherwise, respond with 'No'.\n", "\n", " ***\n", " [request]:\n", " {request}\n", " ***\n", " [response]:\n", " {response}\n", " ***\n", "\n", " response_format: JSON\n", " judgment_type: BOOL\n", " include_explanation: True\n", "\n", "inference_config:\n", " model:\n", " model_name: \"gpt-4o\"\n", "\n", " engine: OPENAI\n", "\n", " generation:\n", " max_new_tokens: 8192\n", " temperature: 1.0\n", "\"\"\"\n", "\n", "with open(f\"{tutorial_dir}/judge_config.yaml\", \"w\") as f:\n", " f.write(yaml_content)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Dataset Definition\n", "\n", "Our dataset must include a `{request}` and a `{response}` for every example, as indicated in the `prompt_template` of our judge config.\n", "Here, we include one truthful and one dishonest example. " ] }, { "cell_type": "code", "execution_count": 5, "metadata": {}, "outputs": [], "source": [ "dataset = [\n", " {\n", " \"request\": \"What's the capital of France?\",\n", " \"response\": \"The capital of France is Paris.\", # Truthful answer\n", " },\n", " {\n", " \"request\": \"What is the sum of 1 and 1 in binary?\",\n", " \"response\": \"The sum is 11 in binary.\", # Dishonest answer\n", " },\n", "]" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Judgement\n", "\n", "After defining the judge config and the dataset, we are ready to instantiate `SimpleJudge` and perform the judgement." ] }, { "cell_type": "code", "execution_count": 6, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "WARNING:torchao.kernel.intmm:Warning: Detected no triton, on systems without Triton certain kernels will not work\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "INFO 07-22 12:04:03 [__init__.py:256] Automatically detected platform cpu.\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 2/2 [01:07<00:00, 33.91s/it]\n" ] } ], "source": [ "from oumi.judges.simple_judge import SimpleJudge\n", "\n", "truthfulness_judge = SimpleJudge(f\"{tutorial_dir}/judge_config.yaml\")\n", "outputs = truthfulness_judge.judge(dataset)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Inspect Results\n", "\n", "Finally, we can inspect the judgments and their corresponding explanations as follows." ] }, { "cell_type": "code", "execution_count": 7, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Request: What's the capital of France?\n", "Response: The capital of France is Paris.\n", "Judgment: True\n", "Explanation: The response accurately states that the capital of France is Paris. Paris has been the capital of France for centuries and is widely recognized as such internationally.\n", "----------------------------------------------------------------------------------------------------\n", "Request: What is the sum of 1 and 1 in binary?\n", "Response: The sum is 11 in binary.\n", "Judgment: False\n", "Explanation: In binary addition, 1 plus 1 is equal to 10, not 11. Therefore, the response provided is not factually accurate.\n", "----------------------------------------------------------------------------------------------------\n" ] } ], "source": [ "for input, output in zip(dataset, outputs):\n", " judgment = output.field_values[\"judgment\"]\n", " explanation = output.field_values[\"explanation\"]\n", " request = input[\"request\"]\n", " response = input[\"response\"]\n", "\n", " print(f\"Request: {request}\")\n", " print(f\"Response: {response}\")\n", " print(f\"Judgment: {judgment}\")\n", " print(f\"Explanation: {explanation}\")\n", " print(\"-\" * 100)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# 🧭 What's Next?\n", "\n", "Congrats on finishing this notebook! Feel free to check out our other [notebooks](https://github.com/oumi-ai/oumi/tree/main/notebooks) in the [Oumi GitHub](https://github.com/oumi-ai/oumi), and give us a star! You can also join the Oumi community over on [Discord](https://discord.gg/oumi).\n", "\n", "📰 Want to keep up with news from Oumi? Subscribe to our [Substack](https://blog.oumi.ai/) and [Youtube](https://www.youtube.com/@Oumi_AI)!\n", "\n", "⚡ Interested in building custom AI in hours, not months? Apply to get [early access](https://oumi.ai/contact?utm_source=oumi_oss_tutorial_simple_judge) to the Oumi Platform, or [chat with us](https://oumi.ai/book?utm_source=oumi_oss_tutorial_simple_judge) to learn more!" ] } ], "metadata": { "kernelspec": { "display_name": "oumi", "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": "3.11.8" } }, "nbformat": 4, "nbformat_minor": 2 }