--- license: apache-2.0 pretty_name: TIGER-Lab/WebInstruct-verified (mirror) source_datasets: - TIGER-Lab/WebInstruct-verified tags: - laya - laya-finetune - mirror - domain-science - domain-general - use-conditional - genre-exam - label-text - annot-natural - lang-en language: - en task_categories: - question-answering size_categories: - 100K 💻 コード | 📄 論文 | 📊 データセット | 🤗 モデル | 🌐 プロジェクトページ

## 概要

General-Reasoner Teaser

図: モデルベースの検証器を使い、多様な検証可能な推論問題で学習した General-Reasoner の、さまざまな推論タスクにおける効果(ベースラインの手法との比較)。

**General-Reasoner** は、大規模言語モデル(LLM)の学習パラダイムであり、数学やコードだけでなく、物理、化学、金融、人文学などを含む多様な分野で推論能力を頑健に高めるように設計されている。 **主な特徴:** - **Zero RL 学習:** 中間の教師あり段階を経ず、ベースの LLM から直接強化学習を行う。 - **多様な推論データ:** Web から集め、分野を越えて答えの検証可能性でフィルタリングした、230K 件以上の高品質で検証可能な問題。 - **モデルベースの検証器:** 文脈を考慮し、思考の連鎖を伴って答えを検証する、1.5B のコンパクトな生成型の検証モデル。従来のルールベースの手法より高い性能を示す。 **本リポジトリには、多様な推論データ WebInstruct-verified が含まれる。** ## データセットの詳細 数学を超えて広い分野で頑健な推論能力を育てるため、多様で高品質なデータセットを構築した。 - **WebInstruct のデータを元の Web ページまで遡り、質問と答えの組を再クロールした。** 元のページに人の手で書かれた答えがない場合は、その項目を捨てた。これにより、再クロールされたすべての項目が人の手で検証済みであり、したがって各答えの品質が信頼できることを保証する。 - **Gemini‑1.5‑Pro を使い、明確に検証可能な短い答えを持つ質問を選択的に抽出する。** これによりデータセットの信頼性がさらに高まる。 - **次に Gemini‑2.0‑Flash が、追加のフィルタリングのために質問ごとに 8 件の候補の答えを生成する:** - Gemini が生成した **8 件の答えがすべて誤っている** 質問は捨てる。これにより、Web スクレイピングの際に生じた曖昧でノイズの多い項目を除く。 - また、**過度に単純な質問**、すなわち **8 件の候補の答えがすべて正しい** 質問も除く。データセットの複雑さを保ち、モデルの汎化をより強く試すためである。 これらの手順により、構築したデータセットの正しさが保証される。 ## 分布 分野の分布は次の図のとおり。 ## 検証 短い答えの形式は、float、array、matrix、latex など多様である。これらの答えを検証するには、GPT/Gemini を使うか、https://huggingface.co/TIGER-Lab/general-verifier でローカルに提供されているモデルを使う。 ## 注記 - [discussion_3](https://huggingface.co/datasets/TIGER-Lab/WebInstruct-verified/discussions/3#6912ee9987d86c668866171a) で @zlk が指摘したとおり、一部の選択式問題では、質問文から選択肢が欠けていた。最新のデータセットではこれを修正した。元のバージョンのデータセットは `train_legacy` split にある。 ## 引用 本研究が役立つと考えられる場合は、次のように引用すること: ```bibtex @inproceedings{ ma2025generalreasoner, title={{G}eneral-{R}easoner: Advancing {LLM} Reasoning Across All Domains}, author={Xueguang Ma and Qian Liu and Dongfu Jiang and Ge Zhang and Zejun MA and Wenhu Chen}, booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}, year={2025}, url={https://openreview.net/forum?id=pBFVoll8Xa} } ```
原文(英語) # General-Reasoner: Advancing LLM Reasoning Across All Domains

💻 Code | 📄 Paper | 📊 Dataset | 🤗 Model | 🌐 Project Page

## Overview

General-Reasoner Teaser

Figure: Effectiveness of General-Reasoner trained with diverse verifiable reasoning questions using model-based verifier compared to baseline methods on various reasoning tasks.

**General-Reasoner** is a training paradigm for large language models (LLMs), designed to robustly enhance reasoning abilities across diverse domains—not just mathematics and coding, but also physics, chemistry, finance, humanities, and more. **Key features:** - **Zero RL Training:** Direct reinforcement learning from base LLMs, bypassing intermediate supervised stages. - **Diverse Reasoning Data:** 230K+ high-quality, verifiable questions sourced from the web and filtered for answer verifiability across disciplines. - **Model-Based Verifier:** Compact 1.5B generative verifier model for context-aware, chain-of-thought answer validation, outperforming traditional rule-based methods. **This repo contains the Diverse Reasoning Data WebInstruct-verified ** ## Dataset Details We construct a diverse, high‑quality dataset to facilitate robust reasoning capabilities across a broad range of domains, extending beyond the commonly studied mathematical problems. - **We trace back the data in WebInstruct to its original web page to re‑crawl the question–answer pairs.** If the original page lacks human‑written answers, we drop the entry. This ensures every re‑crawled item is human‑verified and, therefore, that each answer is of reliable quality. - **Gemini‑1.5‑Pro is employed to selectively extract questions with clearly verifiable short answers,** further boosting dataset reliability. - **Gemini‑2.0‑Flash then generates eight candidate answers per question for additional filtering:** - We discard any question for which **all eight Gemini‑generated answers are incorrect**, eliminating ambiguous or noisy items that arose during web scraping. - We also remove **overly simple questions**—those for which **all eight candidate answers are correct**—to preserve dataset complexity and better challenge model generalization. These steps ensure the correctness of the constructed dataset. ## Distribution The distribution of disciplines is depicted as follows: ## Verification The short answers have different forms, including float, array, matrix, latex, etc. To verify these answers, please use GPT/Gemini or use the locally-served model at https://huggingface.co/TIGER-Lab/general-verifier. ## Notes - As discussed in [discussion_3](https://huggingface.co/datasets/TIGER-Lab/WebInstruct-verified/discussions/3#6912ee9987d86c668866171a), noticed by @zlk, some multiple choice questions have options missing from the question. We have fixed this in the latest dataset. The original version of the dataset can be found in the `train_legacy` split. ## Citation If you feel our work is helpful, please cite: ```bibtex @inproceedings{ ma2025generalreasoner, title={{G}eneral-{R}easoner: Advancing {LLM} Reasoning Across All Domains}, author={Xueguang Ma and Qian Liu and Dongfu Jiang and Ge Zhang and Zejun MA and Wenhu Chen}, booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}, year={2025}, url={https://openreview.net/forum?id=pBFVoll8Xa} } ```