datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
fusion-synth-data-s1kx
Offline Synthetic Data (s1K-X) for: Making, not taking, the Best-of-N
Content
This data contains completions for the s1K-X training split prompts from 5 different teacher models and 2 aggregations:
Teachers: We sample one completion from each of the following models at temperature T=0.3. For kimik2, qwen3, and deepseek-v3 we use TogetherAI, for gemma3-27b and command-a we use locally hosted images.
gemma3-27b: GEMMA3-27B-IT
kimik2: KIMI-K2-INSTRUCT
qwen3: QWEN3-235B… See the full description on the dataset page: https://huggingface.co/datasets/CohereLabs/fusion-synth-data-s1kx.Natural-Reasoning-gpt-oss-120B-S1
Dataset Card: Natural-Reasoning-gpt-oss-120B-S1
📜 Dataset Overview
This is a meticulously curated instruction fine-tuning dataset designed specifically for efficient knowledge distillation tasks. Built upon the first 100,000 questions from the large-scale reasoning corpus facebook/natural_reasoning (s1, I will process the remaining parts later), it aims to transfer the advanced, multi-step reasoning capabilities of the teacher model gpt-oss-120-high to a student model… See the full description on the dataset page: https://huggingface.co/datasets/Jackrong/Natural-Reasoning-gpt-oss-120B-S1.less-is-moe-s1-calibration-128-seq8192
Less-is-MoE S1K calibration data — 128 samples, seq_length 8192
This is the fixed calibration artifact used to prune GPT-OSS-120B,
Qwen3.5-122B-A10B, and the Gemma-4-26B-A4B causal language tower. It uses the same 128 source rows as the full-length variant:
yentinglin/s1K-1.1-trl-format revision
58a01564d278477da20ead1bcf1cde8e31f36251, train, followed by
Dataset.shuffle(seed=1234) and the first 128 nonempty messages rows.
For pruning, concatenate messages[].content with one… See the full description on the dataset page: https://huggingface.co/datasets/jayzou3773/less-is-moe-s1-calibration-128-seq8192.Amazon-Beauty-S1This dataset is derived from Amazon Reviews'23 [1] Beauty category. The split is standard leave-one-out: the last item is the test target, the second-to-last is the validation target, and everything before that is training.
The training portion is expanded by sliding window — every prefix becomes one example — so a user with a sequence of length $L$ contributes $L-3$ training rows with histories of length $1 \dots L-3$, one validation row and one test row.
Statistics… See the full description on the dataset page: https://huggingface.co/datasets/Abner0803/Amazon-Beauty-S1.s1K-1.1-deepseek-cot
s1K-1.1 (DeepSeek-R1 traces) — format cho SegmentSelectiveSFT
Chuyen doi tu simplescaling/s1K-1.1
bang prepare_s1k.py (default flags) trong repo SegmentSelectiveSFT.
Moi dong jsonl co 3 truong:
Truong
Nguon
question
question
solution
deepseek_thinking_trajectory (long-CoT trace cua R1)
answer
\\boxed{...} cuoi cung trong trace, fallback ve solution cua s1K
Giu 934 / 1000 mau — bo cac mau khong co trace, khong co dap an, hoac dap an dai hon 200 ky tu.
from… See the full description on the dataset page: https://huggingface.co/datasets/baesad/s1K-1.1-deepseek-cot.ANGO-S1ANGO is A Novel Generation-Oriented Chinese LLM evaluation benchmark.
We introduces the format of single-question multiple-keypoints dataset for the first time, which include 171 keypoints accumulated in 4 hierarchical levels and 9 difficulty categories.
The data were exclusively obtained from the Administrative Proficiency Test, which serves as a significant component of the Chinese civil service examination.
We will apply a seasonal system for the leaderboard, updating them every two months.… See the full description on the dataset page: https://huggingface.co/datasets/AngoHF/ANGO-S1.s1-test-time-scaling-synth-public
s1-test-time-scaling-synth: Japanese and English Reinforcement Learning Dataset Derived from the s1 Simple Test-Time Scaling Dataset
This repository contains s1-test-time-scaling-synth, a reinforcement learning dataset in Japanese and English.This dataset is built upon the supervised fine-tuning dataset simplescaling/data_ablation_full59K (hereafter, the "original dataset"), originally developed in "s1: Simple test-time scaling" [Muennighoff+, EMNLP25].
The original dataset is a… See the full description on the dataset page: https://huggingface.co/datasets/tokyotech-llm/s1-test-time-scaling-synth-public.less-is-moe-s1-calibration-128
Less-is-MoE S1K calibration data — 128 full-length samples
This repository contains the exact 128 S1K rows selected for Less-is-MoE
full-model pruning. The selection reproduces the released loader:
source: yentinglin/s1K-1.1-trl-format
revision: 58a01564d278477da20ead1bcf1cde8e31f36251
split: train
order: Dataset.shuffle(seed=1234)
samples: first 128 nonempty messages rows
sequence-length limit: none
truncation: disabled
padding: disabled
calibration.jsonl stores every… See the full description on the dataset page: https://huggingface.co/datasets/jayzou3773/less-is-moe-s1-calibration-128.S1_QFFT
📘 S1–QFFT
S1–QFFT is a question-free version of the original simplescaling/s1K-1.1 dataset, designed for QFFT training workflows.
🔍 Description
This dataset discards the original questions and any system instructions, keeping only the reasoning completions as supervision. It is especially useful for models that aim to learn when and how to think, rather than just how to answer.
The dataset is fully converted into a format compatible with LLaMA-Factory training.… See the full description on the dataset page: https://huggingface.co/datasets/lwl-uestc/S1_QFFT.S1.1-QwQ
S1.1-QwQ
This dataset is a refined dataset based on s1.1-1k. Specifically, we used QwQ-32B to generate reasoning trajectories for each problem in s1.1-1k dataset.
Based on this dataset, we finetuned our new model based on Qwen2.5-32B-Instruct and DeepSeek-R1-Distill-32B series, we now open-source our best model.
Model: S1.1-QwQ-DS.
Evaluation Results
Category
Model
Model SizeAIME 2024
AIME 2025
MATH500
GPQA
Industrial Models
QwQ
32B
80.0
60.0
97.6
68.2… See the full description on the dataset page: https://huggingface.co/datasets/BitStarWalkin/S1.1-QwQ.Natural-Reasoning-gpt-oss-120B-S1
Dataset Card: Natural-Reasoning-gpt-oss-120B-S1
📜 Dataset Overview
This is a meticulously curated instruction fine-tuning dataset designed specifically for efficient knowledge distillation tasks. Built upon the first 100,000 questions from the large-scale reasoning corpus facebook/natural_reasoning (s1, I will process the remaining parts later), it aims to transfer the advanced, multi-step reasoning capabilities of the teacher model gpt-oss-120-high to a student model… See the full description on the dataset page: https://huggingface.co/datasets/JAMhunggingface/Natural-Reasoning-gpt-oss-120B-S1.