search
HomeTechnology peripheralsAICMU & Tsinghua's new work: Let LLM synthesize data to learn by itself, and the performance of specific tasks is also greatly improved.

CMU & Tsinghuas new work: Let LLM synthesize data to learn by itself, and the performance of specific tasks is also greatly improved.

The AIxiv column is a column where this site publishes academic and technical content. In the past few years, the AIxiv column of this site has received more than 2,000 reports, covering top laboratories from major universities and companies around the world, effectively promoting academic exchanges and dissemination. If you have excellent work that you want to share, please feel free to contribute or contact us for reporting. Submission email: liyazhou@jiqizhixin.com; zhaoyunfeng@jiqizhixin.com The main authors of this article are from Tsinghua University and Carnegie Mellon University (CMU). Together they are Zhao Chenyang, an undergraduate graduate of the Computer Science Department of Tsinghua University, and Jia Xueying, a master's student of Carnegie Mellon University.
Although large-scale language models (LLM) perform well in many natural language processing tasks, their results in specific tasks are not satisfactory. In order to improve the performance of models on specific natural language tasks, existing methods mainly rely on high-quality manually annotated data. The process of collecting this type of data is time-consuming and laborious, especially difficult for tasks where data are scarce.
In order to solve this problem, some research attempts to generate training data through powerful Teacher Model to enhance the performance of Student Model on specific tasks. However, this approach still faces many challenges in terms of cost, scalability, and legal compliance. When high-quality human supervision signals cannot be continuously obtained, the ability to continuously iterate the model has become an urgent problem to be solved.
A research team from Carnegie Mellon University and Tsinghua University proposed the SELF-GUIDE method. This method generates a task-specific data set by the language model itself and fine-tunes it on this data set, thereby significantly improving the model's ability on a specific task without relying on a large amount of external high-quality data or a more powerful Teacher Model. Specifically, with approximately 3 external input samples, SELF-GUIDE uses a multi-stage generation and filtering mechanism to fine-tune the model using synthetic data generated by the model to make the model perform better on specific tasks.

CMU & Tsinghuas new work: Let LLM synthesize data to learn by itself, and the performance of specific tasks is also greatly improved.

Paper address: https://arxiv.org/abs/2407.12874 Code repository: https://github.com/zhaochenyang20/Prompt2Model-SELF-GUIDE

CMU & Tsinghuas new work: Let LLM synthesize data to learn by itself, and the performance of specific tasks is also greatly improved.

                     Figure 1: SELF-GUIDE uses the model to autonomously synthesize data The ability to improve the model's ability to perform specific tasks.

Method

Specifically, the research team decomposed the SELF-GUIDE method into three main stages: input data generation, output data generation and quality optimization.

Input data generation

In the design and implementation process of the SELF-GUIDE framework, the researcher first specified different prompt templates according to the task type (generative task or classification task). For generative tasks, the SELF-GUIDE framework uses a relatively simple prompt template. For classification tasks, the SELF-GUIDE framework adopts another strategy. For classification tasks, the SELF-GUIDE framework first randomly selects a label from all label spaces and uses it as a conditionally generated pseudo-label to guide the generation of input data. After selecting a pseudo-label, the SELF-GUIDE framework uses more complex conditions to generate a template to guide the model to generate input content corresponding to the selected pseudo-label.

CMU & Tsinghuas new work: Let LLM synthesize data to learn by itself, and the performance of specific tasks is also greatly improved.

Figure 2: The core of SELF-GUIDE lies in an efficient multi-stage generation mechanism, in which the language model gradually generates input-output data combinations. After generation and filtering, the self-generated data is further used to fine-tune the language model itself. This diagram describes SELF-GUIDE's flow for build tasks.

After the template is selected and the few-shot examples are populated, the complete prompt is passed to LLM to generate the input data. After each round of prompts, newly generated inputs are added to the input library. A subset of the inputs are randomly sampled from this library and merged with the inputs from the initial example to form new cues, gradually expanding the set of inputs generated by the LLM and reducing duplication. SELF-GUIDE performs only one round of input generation, followed by a quality optimization phase where rule-based filters are applied to remove low-quality inputs.

CMU & Tsinghuas new work: Let LLM synthesize data to learn by itself, and the performance of specific tasks is also greatly improved.


図 3: この図は、SELF-GUIDE が分類タスクを完了するプロセスを示しています。分類タスクからのデータの場合、SELF-GUIDE は最初に疑似ラベルを生成し、次に対応する入力を生成し、最後に実際のラベルを再生成します。
出力データ生成
出力データ生成フェーズでは、典型的なコンテキスト学習方法が使用されます。研究者はタスクの指示と元の例をモデルに提供し、モデルが入力生成フェーズで生成された各入力にラベルを付けることができるようにします。すべての出力が取得された後、別のルールベースのフィルタリングが実行されて、最終的な合成データセットが選択されます。
品質の最適化
生成されたデータの品質は、下流のトレーニングの成功にとって重要です。 SELF-GUIDE では、品質を向上させるために 2 つの戦略を採用しています。1 つは生成パラメータを調整して生成品質を向上させる方法、もう 1 つはルールに基づいて低品質のサンプルをフィルタリングする方法です。
温度を調整する: 温度を調整することは、品種と品質のバランスをとるための一般的な戦略です。 SELF-GUIDE フレームワークは、入力生成ステージでより高い温度を使用して多様性を促進し、他のステージでより低い温度を使用して最高確率の出力を保証し、全体的なデータ品質を保証します。ただし、温度調整だけでは望ましいバランスを達成するのに十分ではありません。したがって、SELF-GUIDE は、入力生成後と出力アノテーション後に 2 回のルールベースのデータ フィルタリングも実行します。
ノイズ フィルター: 研究者は、一般的な挨拶やノイズ文字 (たとえば、生成されたコンテンツ内の「”」など) を含むノイズ用語のリストを手動で作成しました。生成されたノイズ用語の例の入力または出力に含まれます。
長さフィルター: サンプルの長さは偏っている可能性がありますが、研究者は、これらのサンプルが特定のタスクの長さの分布の観点からは代表的なものであると信じています。サンプルの長さは正規分布に従い、入力サンプルの平均 μ と標準偏差 σ を計算します。研究者は、生成されたサンプルの入力長と出力長が同じ正規分布に従う必要があると想定し、長さを必要とします。 (μ − 2σ、μ + 2σ)
1 つのパラメーターですべてに適合: SELF-GUIDE が指示と例で指定されたターゲット分布に準拠するトレーニング データを生成するには、ラベル付けされたパラメータでさまざまなハイパーパラメーターを最適化する必要があります。生成された入力と出力の数、入力データが生成される温度、出力データが生成される温度、パラメータの微調整などのデータ ポイント。研究者は実験テスト タスクを 2 つの部分に分割します。生成パラメータを調整するためにすべてのデータを使用できます。これは検証タスクと呼ばれます。データの他の部分はテストにのみ使用され、研究者が検索するパラメータの調整には使用できません。検証タスクで「最悪のタスクのパフォーマンスを最大化」するパラメータを設定し、SELF-GUIDE のパフォーマンスを評価するために修正しました。研究者は、Super-NaturalInstructions V2 ベンチマークからタスクの半分をランダムに選択し、残りの半分を評価に使用しました。入力生成、出力生成、微調整については、研究者は Super-Natural 命令ベンチマークと同じものを使用し、分類タスクには完全一致、生成タスクには ROUGE-L を使用しました。 SELF-GUIDE の効果を反映するために、研究者は SELF-GUIDE を他の命令追従およびコンテキスト学習方法と比較しました。
1.Few-Shot ICL: 主なベンチマークとして、研究者は直接ヒント言語モデルと比較しました。
2. Self-ICL は、自己生成されたサンプルを使用して、Self-ICL の作業に基づいて修正を加え、命令追従の数を増やします。プロンプトの単語を埋めるために (固定数の例ではなく) できるだけ多くの例を自己生成することで、サンプルを参照します。
3.少数ショット微調整: 微調整には少数の入力サンプルを直接使用します。
セルフガイド 原文の主な実験結果は以下の通りです。ベースライン評価指標では、分類タスクの絶対改善は 14.5% に達し、生成タスクの絶対改善は 17.9% に達しました。これらの結果は、データが非常に限られている場合でも、SELF-GUIDE が LLM をタスク固有の専門化に導くのに非常に効果的であることを示しています。これは、LLM を特定のタスクに大規模に適応させるための自己生成データの可能性を強調しています。より詳しい実験結果やアブレーション実験については、原論文を参照してください。



図 4: タスクのタイプ (分類タスクと生成タスク) ごとに、研究者はタスクをランダムに 2 つの半分に分割し、半分は「1 つのパラメーターがすべてに適合」戦略のパラメーターのデバッグに使用され、もう半分は使用されました。半分は、これらのデバッグされたパラメーターを使用するために使用されました。パラメーターは、SELF-GUIDE のパフォーマンスをテストします。 SELF-GUIDE の前後でモデルのパフォーマンスを評価するために、同じデコード パラメーターとキュー テンプレートを使用します。

CMU & Tsinghuas new work: Let LLM synthesize data to learn by itself, and the performance of specific tasks is also greatly improved.概要

SELF-GUIDE フレームワークは、モデルが自律的にトレーニング データを生成し、このデータに基づいて微調整することを促進します。実験結果は、この方法が特定のタスクに対する大規模言語モデルの専門的能力を向上させる大きな可能性を秘めていることを示しています。特にデータが限られている場合、SELF-GUIDE はトレーニング データの不足の問題を効果的に解決できます。同時に、これは自律的なモデル適応と継続的学習のためのテクノロジーを探索するための参考にもなります。研究者らは、この研究が自律的な調整と改善メカニズムの AI システムの開発を促進し、AI システムを人間の意図とより一致させることを期待しています。

The above is the detailed content of CMU & Tsinghua's new work: Let LLM synthesize data to learn by itself, and the performance of specific tasks is also greatly improved.. For more information, please follow other related articles on the PHP Chinese website!

Statement
The content of this article is voluntarily contributed by netizens, and the copyright belongs to the original author. This site does not assume corresponding legal responsibility. If you find any content suspected of plagiarism or infringement, please contact admin@php.cn
[Ghibli-style images with AI] Introducing how to create free images with ChatGPT and copyright[Ghibli-style images with AI] Introducing how to create free images with ChatGPT and copyrightMay 13, 2025 am 01:57 AM

The latest model GPT-4o released by OpenAI not only can generate text, but also has image generation functions, which has attracted widespread attention. The most eye-catching feature is the generation of "Ghibli-style illustrations". Simply upload the photo to ChatGPT and give simple instructions to generate a dreamy image like a work in Studio Ghibli. This article will explain in detail the actual operation process, the effect experience, as well as the errors and copyright issues that need to be paid attention to. For details of the latest model "o3" released by OpenAI, please click here⬇️ Detailed explanation of OpenAI o3 (ChatGPT o3): Features, pricing system and o4-mini introduction Please click here for the English version of Ghibli-style article⬇️ Create Ji with ChatGPT

Explaining examples of use and implementation of ChatGPT in local governments! Also introduces banned local governmentsExplaining examples of use and implementation of ChatGPT in local governments! Also introduces banned local governmentsMay 13, 2025 am 01:53 AM

As a new communication method, the use and introduction of ChatGPT in local governments is attracting attention. While this trend is progressing in a wide range of areas, some local governments have declined to use ChatGPT. In this article, we will introduce examples of ChatGPT implementation in local governments. We will explore how we are achieving quality and efficiency improvements in local government services through a variety of reform examples, including supporting document creation and dialogue with citizens. Not only local government officials who aim to reduce staff workload and improve convenience for citizens, but also all interested in advanced use cases.

What is the Fukatsu-style prompt in ChatGPT? A thorough explanation with example sentences!What is the Fukatsu-style prompt in ChatGPT? A thorough explanation with example sentences!May 13, 2025 am 01:52 AM

Have you heard of a framework called the "Fukatsu Prompt System"? Language models such as ChatGPT are extremely excellent, but appropriate prompts are essential to maximize their potential. Fukatsu prompts are one of the most popular prompt techniques designed to improve output accuracy. This article explains the principles and characteristics of Fukatsu-style prompts, including specific usage methods and examples. Furthermore, we have introduced other well-known prompt templates and useful techniques for prompt design, so based on these, we will introduce C.

What is ChatGPT Search? Explains the main functions, usage, and fee structure!What is ChatGPT Search? Explains the main functions, usage, and fee structure!May 13, 2025 am 01:51 AM

ChatGPT Search: Get the latest information efficiently with an innovative AI search engine! In this article, we will thoroughly explain the new ChatGPT feature "ChatGPT Search," provided by OpenAI. Let's take a closer look at the features, usage, and how this tool can help you improve your information collection efficiency with reliable answers based on real-time web information and intuitive ease of use. ChatGPT Search provides a conversational interactive search experience that answers user questions in a comfortable, hidden environment that hides advertisements

An easy-to-understand explanation of how to create a composition in ChatGPT and prompts!An easy-to-understand explanation of how to create a composition in ChatGPT and prompts!May 13, 2025 am 01:50 AM

In a modern society with information explosion, it is not easy to create compelling articles. How to use creativity to write articles that attract readers within a limited time and energy requires superb skills and rich experience. At this time, as a revolutionary writing aid, ChatGPT attracted much attention. ChatGPT uses huge data to train language generation models to generate natural, smooth and refined articles. This article will introduce how to effectively use ChatGPT and efficiently create high-quality articles. We will gradually explain the writing process of using ChatGPT, and combine specific cases to elaborate on its advantages and disadvantages, applicable scenarios, and safe use precautions. ChatGPT will be a writer to overcome various obstacles,

How to create diagrams using ChatGPT! Illustrated loading and plugins are also explainedHow to create diagrams using ChatGPT! Illustrated loading and plugins are also explainedMay 13, 2025 am 01:49 AM

An efficient guide to creating charts using AI Visual materials are essential to effectively conveying information, but creating it takes a lot of time and effort. However, the chart creation process is changing dramatically due to the rise of AI technologies such as ChatGPT and DALL-E 3. This article provides detailed explanations on efficient and attractive diagram creation methods using these cutting-edge tools. It covers everything from ideas to completion, and includes a wealth of information useful for creating diagrams, from specific steps, tips, plugins and APIs that can be used, and how to use the image generation AI "DALL-E 3."

An easy-to-understand explanation of ChatGPT Plus' pricing structure and payment methods!An easy-to-understand explanation of ChatGPT Plus' pricing structure and payment methods!May 13, 2025 am 01:48 AM

Unlock ChatGPT Plus: Fees, Payment Methods and Upgrade Guide ChatGPT, a world-renowned generative AI, has been widely used in daily life and business fields. Although ChatGPT is basically free, the paid version of ChatGPT Plus provides a variety of value-added services, such as plug-ins, image recognition, etc., which significantly improves work efficiency. This article will explain in detail the charging standards, payment methods and upgrade processes of ChatGPT Plus. For details of OpenAI's latest image generation technology "GPT-4o image generation" please click: Detailed explanation of GPT-4o image generation: usage methods, prompt word examples, commercial applications and differences from other AIs Table of contents ChatGPT Plus Fees Ch

Explaining how to create a design using ChatGPT! We also introduce examples of use and promptsExplaining how to create a design using ChatGPT! We also introduce examples of use and promptsMay 13, 2025 am 01:47 AM

How to use ChatGPT to streamline your design work and increase creativity This article will explain in detail how to create a design using ChatGPT. We will introduce examples of using ChatGPT in various design fields, such as ideas, text generation, and web design. We will also introduce points that will help you improve the efficiency and quality of a variety of creative work, such as graphic design, illustration, and logo design. Please take a look at how AI can greatly expand your design possibilities. table of contents ChatGPT: A powerful tool for design creation

See all articles

Hot AI Tools

Undresser.AI Undress

Undresser.AI Undress

AI-powered app for creating realistic nude photos

AI Clothes Remover

AI Clothes Remover

Online AI tool for removing clothes from photos.

Undress AI Tool

Undress AI Tool

Undress images for free

Clothoff.io

Clothoff.io

AI clothes remover

Video Face Swap

Video Face Swap

Swap faces in any video effortlessly with our completely free AI face swap tool!

Hot Article

Hot Tools

Dreamweaver Mac version

Dreamweaver Mac version

Visual web development tools

SublimeText3 Mac version

SublimeText3 Mac version

God-level code editing software (SublimeText3)

EditPlus Chinese cracked version

EditPlus Chinese cracked version

Small size, syntax highlighting, does not support code prompt function

MinGW - Minimalist GNU for Windows

MinGW - Minimalist GNU for Windows

This project is in the process of being migrated to osdn.net/projects/mingw, you can continue to follow us there. MinGW: A native Windows port of the GNU Compiler Collection (GCC), freely distributable import libraries and header files for building native Windows applications; includes extensions to the MSVC runtime to support C99 functionality. All MinGW software can run on 64-bit Windows platforms.

SecLists

SecLists

SecLists is the ultimate security tester's companion. It is a collection of various types of lists that are frequently used during security assessments, all in one place. SecLists helps make security testing more efficient and productive by conveniently providing all the lists a security tester might need. List types include usernames, passwords, URLs, fuzzing payloads, sensitive data patterns, web shells, and more. The tester can simply pull this repository onto a new test machine and he will have access to every type of list he needs.