Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
Job manager crashed while running this job (missing heartbeats).
Error code:   JobManagerCrashedError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

conversations
list
[ { "from": "user", "value": "Find all prime numbers $p$ and positive integers $m$ such that $2p^2 + p + 9 = m^2.$" }, { "from": "assistant", "value": "<think>\nOkay, so I need to find all prime numbers p and positive integers m such that the equation 2p² + p + 9 = m² holds. Hmm, let's start by un...
[ { "from": "user", "value": "27. Find the digit that has the property that if it is appended to the end of any natural number, the resulting number is equal to the sum of three terms, the first of which is the original number, the second is the number represented by the sought digit, and the third is the pro...
[ { "from": "user", "value": "1. In one lyceum, $76 \\%$ of the students have at least once not done their homework, and $\\frac{5}{37}$ sometimes forget their second pair of shoes. Find the number of students in the lyceum, if it is more than 1000 but less than 2000." }, { "from": "assistant", "v...
[ { "from": "user", "value": "\nProblem 3. Find the smallest positive integer $n$, such that there exist $n$ integers $x_{1}, x_{2}, \\ldots, x_{n}$ (not necessarily different), with $1 \\leq x_{k} \\leq n, 1 \\leq k \\leq n$, and such that\n\n$$\nx_{1}+x_{2}+\\cdots+x_{n}=\\frac{n(n+1)}{2}, \\quad \\text { a...
[ { "from": "user", "value": "If we divide a square into parts with three lines, then among the resulting polygons, there is always one whose diameter is not less than $\\sqrt{13}$. What is the minimum length of the side of the square? (The diameter of a polygon is the distance between its two farthest vertic...
[ { "from": "user", "value": "Example 7 Real numbers $a$ and $b$ satisfy the equation $\\frac{a^{2} b^{2}}{a^{4}-2 b^{4}}=$\n1. Then $\\frac{a^{2}-b^{2}}{19 a^{2}+96 b^{2}}=$ $\\qquad$\n(1996, Beijing Junior High School Mathematics Competition)" }, { "from": "assistant", "value": "<think>\nOkay, l...
[ { "from": "user", "value": "10. (12 points) 1 kilogram of soybeans can be made into 3 kilograms of tofu, and 1 kilogram of soybean oil requires 6 kilograms of soybeans. Tofu sells for 3 yuan per kilogram, and soybean oil sells for 15 yuan per kilogram. A batch of soybeans weighs 460 kilograms, and after bei...
[ { "from": "user", "value": "The area of the smallest equilateral triangle with one vertex on each of the sides of the right triangle with side lengths $2\\sqrt{3},~5,$ and $\\sqrt{37},$ as shown, is $\\frac{m\\sqrt{p}}{n},$ where $m,~n,$ and $p$ are positive integers, $m$ and $n$ are relatively prime, and $...
[ { "from": "user", "value": "G10.2 The medians $A L, B M, C N$ of $\\triangle A B C$ meet at $G$. If the area of $\\triangle A B C$ is $54 \\mathrm{~cm}^{2}$ and the area of $\\triangle A N G$ is $x \\mathrm{~cm}^{2}$. Find $x$." }, { "from": "assistant", "value": "<think>\nOkay, so I need to fin...
[ { "from": "user", "value": "12. Let $f(x)=|x-p|+|x-15|+|x-p-15|$, where $0<p<15$, if $p \\leqslant x \\leqslant 15$, then the minimum value of $f(x)$ is $\\qquad$" }, { "from": "assistant", "value": "<think>\nOkay, so I need to find the minimum value of the function f(x) = |x - p| + |x - 15| + |...
[ { "from": "user", "value": "10. Among the first 100 positive integers, the sum of numbers that can be expressed as the difference of squares of two positive integers is" }, { "from": "assistant", "value": "<think>\nOkay, let me try to figure out this problem. So, the question is asking: Among th...
End of preview.

qihoo360/Light-R1-SFTData(複製)

データの内容

数学の問題と推論の会話からなる SFT 用のデータで、Light-R1-32B の 2 段階の学習のうち stage1 と stage2 にあたる。1 行が 1 件の会話で、conversations に問題と応答が入る。汚染除去した数学データを使い、DeepSeek-R1 を蒸留する方法で作られている。数学だけのデータで、難易度の連続値は公開されていない。

データの概要

項目 値
行数 79,439(1 構成・1 分割の合計)
列数 1
容量 1.53 GB(ファイル 4 本)
ファイルの形式 json 2
公開 public

構成と分割

構成(config) 分割(split)ごとの行数 列数
default train 79,439 1

列(構成 default・分割 train)

列 型 先頭の行の値
conversations list<struct{from: string, value: string}> [2 件]

行数と列は HF のデータセットビューアの集計(元のデータセットの現在の版で数えたもの)。全データセットの一覧は HYUKI5/dataset-catalog にある。

複製について

qihoo360/Light-R1-SFTData をリビジョン f939e79f82f0c7c14cbb871ee0b75f8f7226ad73 の時点でそのまま複製したもの。中身は変えていない。判断モデル Laya(convaiinnovations/laya)の微調整に使う候補として、調査メモ(Laya 学習用データセット調査)に載せたデータを集めている。

出典

論文

アノテーションの方法

元のカードにアノテーションの節が無かった。下の「調査メモの記述」と「元のカード」を参照。

微調整に使うときの注意

  • 調査メモの台帳での区分: B(継承条件・gated・モデル由来の条件・上流の規約などが付く)。区分は記載の文言で機械的に付けたもので、法的な判定ではない。学習に入れる前に元のカードとライセンスの原文で確かめる。
  • Laya の入力は 2,048 token まで(ModernBERT-large / mmBERT)。長い本文は先頭から黙って切られるので、切れる行は事前に除くか、要約・切り出しをしてから入れる。
  • 分割は記事・文書単位で行い、同じ本文から作った別の問いを学習側と較正側に分けない。温度と閾値は学習に使っていない分割で別々に当てる。

調査メモの記述

メモ 節 ライセンス(記載) 確認 注記
open_weight_reasoning_logs.md Q3. 難易度ラベル・ドメイン列・検証可能な短答を持つ質問データ(トレースが無い、または質問側に価値があるもの)は何があるか (L88) apache-2.0 確認済 生成元モデルの出力規約が別に及ぶ(Q4) / 原文注記: 数学のみ。難易度の連続値は公開されていない(絞る前提の情報のみ)

元のカード(日本語訳)

以下は元のデータセットのカードの本文を日本語に訳したもの(機械翻訳)。訳に疑問があるときは、末尾の原文を確かめる。


Light-R1: カリキュラム SFT & DPO により、$1000 でゼロから R1-Distill を超える*

*長い COT を持たないモデルから出発した場合

技術報告

GitHub のページ

以下は Light-R1-32B の学習に使った 2 段階の SFT データである。 stage1-76k.json と stage2-3k.json を参照すればよい。

モデル 学習元 公開日 AIME24 AIME25
DeepSeek-R1-Distill-Llama-70B Llama-3.3-70B-Instruct 25.1.20 70.0 54.1
DeepSeek-R1-Distill-Qwen-32B Qwen2.5-32B 25.1.20 72.6 54.9
LIMO (32B) Qwen2.5-32B-Instruct 25.2.4 56.3 47.1
s1.1-32B Qwen2.5-32B-Instruct 25.2.8 64.7 47.8
OpenThinker-32B Qwen2.5-32B-Instruct 25.2.12 66.0 50.9
Light-R1-32B(本研究) 🤗 Qwen2.5-32B-Instruct 25.3.4 76.6 64.6

72B 以下のモデルで DeepSeek-R1 の再現を試みたオープンソースの研究は多いが、難しい数学コンテスト AIME24 で DeepSeek-R1-Distill-Qwen-32B のスコア 72.6 に近い性能を出したものは 一つもない。

Light-R1-32B を紹介する。これは Qwen2.5-32B-Instruct から学習して AIME24 で 76.6 を達成したモデルである。長い COT を持たないモデルから(R1 を基準にすると ゼロから)出発し、汚染除去した数学データで学習した。カリキュラム SFT & DPO により DeepSeek-R1 を蒸留し、AIME24 と 25 で DeepSeek-R1-Distill-Qwen-32B を上回る。さらにモデルのマージで改善した。

さらに重要なこととして、ゼロから学習したモデルとして最先端の Light-R1-32B に加え、カリキュラム SFT & DPO の学習データすべてと、360-LLaMA-Factory に基づく学習コードを初日に公開した。12 台の H800 での学習時間は 6 時間以内と見積もられ、費用は約 $1000 である。

Light-R1 は、強力な長い COT モデルをゼロから(長い COT を持たないモデルから)学習する実用的な方法だと考えている。RL による改善も進めているが、カリキュラム SFT & DPO は学習の流れ全体をより細かく制御でき、コストも抑えられる。

学習と推論の技術が急速に発展する中、近い将来により入手しやすい長い COT のモデルが現れることを期待している。Light-R1 は、少なくとも専門分野においてそれを学習させる、検証済みで透明な方法を示している。

WeChat グループはこちら。

公開の詳細

  • 🤗 huggingface 上の Light-R1-32B モデル

  • カリキュラム 🤗SFT と 🤗DPO データセット

  • 360-LLaMA-Factory に基づく学習スクリプトは train-scripts にある

  • DeepScaleR に基づく評価コードは deepscaler-release にある

    • Light-R1-32B の評価ログ(例: AIME24)も含む
    • 報告しているスコアはすべて 64 回の平均である。公開モデルのスコアは、その評価結果から取り、なければ 64 回の平均を取った。16 回の平均では、実行ごとに 2~3 点ずれることがあるとわかった
  • 技術報告は作成中

推論の注意点

Light-R1-32B は必ずしも思考するとは限らない。思考能力は数学データだけで学習させたためである。

DeepSeek の提案に従い、モデルが出力を生成する直前のチャットテンプレートに <think> を直接書き込んで、Light-R1 に強制的に思考させた。

推論には vLLM または SGLang が推奨される。 Light-R1-32B は Qwen モデルのチャットテンプレートを引き継いでいる。<think> と </think> が特殊トークンとして追加され、思考を強制するため <think> が直接書き込まれている。

カリキュラム SFT & DPO による事後学習

AIME24 pass@1(64 回平均) AIME25 GPQA Diamond
Qwen2.5-32B-Instruct 16.6 13.6 48.8
DeepSeek-R1-Distill-Qwen-32B 72.6 54.9 62.1
Light-R1-SFT-stage1 69.0 57.4 64.3
Light-R1-SFT-stage2 73.0 64.3 60.6
Light-R1-DPO 75.8 63.4 61.8
Light-R1-32B 76.6 64.6 61.8

SFT と DPO によるカリキュラム学習の手法を採用した。

数学データの出典

学習に使った問題は、OpenR1-Math-220k、OpenThoughts-114k、LIMO、OpenMathInstruct-2、s1K-1.1、Omni-MATH、hendrycks_math など、そして 2023 年までの AIME を含む公開の数学データセットから集めた。 AIME24/25、MATH-500、GPQA Diamond など、一般的な推論ベンチマークに対して問題の汚染を除去した。

カリキュラム SFT & DPO

これらの問題に対する DeepSeek-R1 の回答を集め、検証結果と、DeepScaleR-1.5B-Preview をサンプリングして付けた難易度に基づいて絞り込んだ。その結果、SFT ステージ 1 用の 76k データセットができた。

SFT ステージ 1 の後、主に 76k データセットから絞り込んだ、より難しい 3k のデータを SFT ステージ 2 用に作った。

このステージ 2 のデータにより、AIME 24/25 で DeepSeek-R1-Distill-Qwen-32B が 72.6/54.9 から 0.779/0.675 に向上した。

次に、SFT ステージ 2 の後の Light-R1-SFT-stage2 の回答をサンプリングし、問題ごとに正解と不正解を分けた。検証結果と DeepSeek-R1 の回答に基づいて DPO のペアを作った。

DPO(または NCA)は、360-LLaMA-Factory のシーケンス並列化を使って SFT ステージ 2 の上で実行した。

上記の学習工程は比較的速く、12 台の H800 で 6 時間以内に終わると見積もられる。$1000 という見積もりはこのためである。

モデルのマージ

最後に、SFT ステージ 2 のモデル、DPO のモデル、AIME24 のスコアが 74.7 のもう一つの DPO 版のモデルをマージした。 2 つの DPO 版の違いは、一方のデータで不採用の回答の特殊トークンを飛ばしている点である。興味深いことに、その結果のモデルも改善を示した。

この手法では段階的な改善が見られ、各段階の途中の評価結果は上の表に示している。 学習にはまったく使っていない科学の問題の GPQA 評価では、数学に特化した学習によってある程度の忘却が起きた。それでも Light-R1-32B は強い汎化能力を示している。

データの汚染除去

公開されているいくつかのデータセットについて、データの汚染を慎重に評価した。 事前学習では汚染が避けられない場合もあるが、事後学習でベンチマークと比較するのは許容できない。 MATH-500 は、数十問が同一か数字だけが変わった問題になっており、いくらか損なわれている。AIME 24 と 25 は無傷だが、2023 年までの AIME データを取り込む際は特に注意が必要である。

Light-R1-32B では、完全一致と N-gram 一致による徹底した汚染除去を行った。

ライセンスと謝辞

本プロジェクトの公開物はすべてオープンソースライセンスの Apache 2.0 に従う。

学習実験は 360-LLaMA-Factory を用いている。 評価スクリプトは DeepScaleR に基づき、したがって verl にも基づく。

Light-R1-32B は Qwen2.5-32B-Instruct から学習された。 学習データは様々な公開元から集めた。

引用

@misc{lightr1proj,
      title={Light-R1: Curriculum SFT, DPO and RL for Long COT from Scratch and Beyond}, 
      author={Liang Wen, Fenrui Xiao, Xin He, Yunke Cai, Qi An, Zhenyu Duan, Yimin Du, Junchen Liu, Lifu Tang, Xiaowei Lv, Haosheng Zou, Yongchao Deng, Shousheng Jia, Xiangzheng Zhang},
      year={2025},
      eprint={2503.10460},
      archivePrefix={arXiv},
      url={https://github.com/Qihoo360/Light-R1}, 
}
原文(英語)

Light-R1: Surpassing R1-Distill from Scratch* with $1000 through Curriculum SFT & DPO

*from models without long COT

technical report

GitHub page

Here are the two-stage SFT data we used to train Light-R1-32B. Simply refer to stage1-76k.json and stage2-3k.json

Model Trained From Release Date AIME24 AIME25
DeepSeek-R1-Distill-Llama-70B Llama-3.3-70B-Instruct 25.1.20 70.0 54.1
DeepSeek-R1-Distill-Qwen-32B Qwen2.5-32B 25.1.20 72.6 54.9
LIMO (32B) Qwen2.5-32B-Instruct 25.2.4 56.3 47.1
s1.1-32B Qwen2.5-32B-Instruct 25.2.8 64.7 47.8
OpenThinker-32B Qwen2.5-32B-Instruct 25.2.12 66.0 50.9
Light-R1-32B (ours) 🤗 Qwen2.5-32B-Instruct 25.3.4 76.6 64.6

While much work has been open-sourced trying to reproduce DeepSeek-R1 on models of 72B or less, none achieves similar performance on the hard math competition AIME24 as DeepSeek-R1-Distill-Qwen-32B's score 72.6.

We introduce Light-R1-32B, which achieves 76.6 on AIME24 training from Qwen2.5-32B-Instruct. Starting from models without long COT (from scratch in terms of R1) and training on decontaminated math data, we distilled DeepSeek-R1 with curriculum SFT & DPO to surpass DeepSeek-R1-Distill-Qwen-32B on AIME24 & 25, and improved further with model merging.

More importantly, besides the state-of-the-art from-scratch model Light-R1-32B, we also released on Day 1 all training datasets of our curriculum SFT & DPO and training code based on 360-LLaMA-Factory. Estimated training time on 12 x H800 machines takes no more than 6 hours --- around $1000.

We believe Light-R1 represents a practical way of training strong long COT models from scratch (from models without long COT). While we are working to further improve our models with RL, curriculum SFT & DPO facilitates more control along the pipeline and is more cost-friendly.

With the rapid development of training and inference techniques, we hope to see more accessible long-COT models in the near future and Light-R1 provides a validated transparent way to train them in at least specialized domains.

WeChat Group here.

Release Details

  • Light-R1-32B model on 🤗 huggingface

  • Curriculum 🤗SFT & 🤗DPO datasets

  • Training scripts based on 360-LLaMA-Factory in train-scripts

  • Evaluation code based on DeepScaleR in deepscaler-release

    • along with evaluation logs of Light-R1-32B (e.g. AIME24)
    • all our reported scores are averaged over 64 runs; public models' scores are taken from their evaluation results and if not present, averaged over 64 runs; we found that averaging over 16 runs sometimes leads to deviation over 2-3 points across different runs
  • Technical report work in progress

Inference Notes

Light-R1-32B does not always think as its thinking capabilities are trained only with math data.

We forced Light-R1 to think by hard-coding <think> in the chat template right before the model is supposed to generate output, as suggested by DeepSeek.

vLLM or SGLang are suggested for inference. Light-R1-32B inherits Qwen models' chat template with <think> and </think> added as special tokens and <think> hard-coded to force thinking.

Post-Training through Curriculum SFT & DPO

AIME24 pass@1 (64 average) AIME25 GPQA Diamond
Qwen2.5-32B-Instruct 16.6 13.6 48.8
DeepSeek-R1-Distill-Qwen-32B 72.6 54.9 62.1
Light-R1-SFT-stage1 69.0 57.4 64.3
Light-R1-SFT-stage2 73.0 64.3 60.6
Light-R1-DPO 75.8 63.4 61.8
Light-R1-32B 76.6 64.6 61.8

We adopted a curriculum learning approach with SFT and DPO.

Math Data Sources

Training questions are collected from public math datasets including OpenR1-Math-220k, OpenThoughts-114k, LIMO, OpenMathInstruct-2, s1K-1.1, Omni-MATH, hendrycks_math and AIME (up to 2023). We decontaminated the questions against common Reasoning benchmarks such as AIME24/25, MATH-500 and GPQA Diamond.

Curriculum SFT & DPO

We collected responses from DeepSeek-R1 on these questions and filtered them based on verification and difficulty levels rated by sampling DeepScaleR-1.5B-Preview, forming a 76k dataset for SFT stage1.

After SFT stage1, a more difficult set, mostly filtered from the 76k dataset, was constructed with 3k data for SFT stage2.

This stage2 data could boost DeepSeek-R1-Distill-Qwen-32B from 72.6/54.9 to 0.779/0.675 on AIME 24/25.

Then we sampled Light-R1-SFT-stage2's responses after SFT stage2, filtered correct and incorrect ones for each question and construct DPO pairs based on verification results and DeepSeek-R1's responses.

DPO(or NCA) is performed on top of SFT stage2 with sequence parallelism in 360-LLaMA-Factory.

The above training steps are fairly fast and are estimated to finish in less than 6 hours on 12 x H800 machines, hence the estimate of $1000.

Model Merging

Finally, we merged models of SFT-stage2, DPO and another DPO version with AIME24 score 74.7. The two DPO versions differ in that one of the data has special tokens skipped in rejected responses. Interestingly, the resulting version also exhibits improvement.

We observed stepwise improvement in our approach and intermediate evaluation results of each stage are listed in the table above. On the GPQA evaluation of scientific questions we didn't train on at all, math-specialized training has led to some degree of forgetting. However, Light-R1-32B still demonstrates strong generalization ability.

Data Decontamination

We carefully evaluated data contamination of several open-sourced datasets. While certain contamination may be inevitable during pre-training, it is unacceptable for post-training to compare on benchmarks. MATH-500 is somewhat compromised with tens of questions that are identical or only numbers changed. AIME 24 and 25 stay intact but we have to pay special attention when we incorporate AIME data up to 2023.

Light-R1-32B did thorough decontamination with exact or N-gram matching.

License & Acknowledgements

All released materials of this project follow the open-source license Apache 2.0.

Our training experiments are powered by 360-LLaMA-Factory. Our evaluation scripts are based on DeepScaleR and therefore verl.

Light-R1-32B is trained from Qwen2.5-32B-Instruct. Training data are collected from various public sources.

Citation

@misc{lightr1proj,
      title={Light-R1: Curriculum SFT, DPO and RL for Long COT from Scratch and Beyond}, 
      author={Liang Wen, Fenrui Xiao, Xin He, Yunke Cai, Qi An, Zhenyu Duan, Yimin Du, Junchen Liu, Lifu Tang, Xiaowei Lv, Haosheng Zou, Yongchao Deng, Shousheng Jia, Xiangzheng Zhang},
      year={2025},
      eprint={2503.10460},
      archivePrefix={arXiv},
      url={https://github.com/Qihoo360/Light-R1}, 
}
Downloads last month
45

Paper for HYUKI5/qihoo360__Light-R1-SFTData