datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
DRIM-ColdstartSFTThis repository contains the cold-start SFT datasets used in the paper Deep But Reliable: Advancing Multi-turn Reasoning for Thinking with Images
Spec-o3-ColdStartSFT
Spec-o3 Cold-Start (iMCoT) Dataset
Project Page | Paper | Code
Dataset Summary
This dataset contains expert-approved spectral inspection trajectories for cold-start supervised fine-tuning (SFT) of Spec-o3, a tool-augmented vision-language agent for astronomer-aligned spectral inspection and candidate vetting.
Each sample is an interleaved multimodal chain-of-thought (iMCoT) trajectory that alternates between:
Textual inspection reasoning, and
Structured tool calls that… See the full description on the dataset page: https://huggingface.co/datasets/Maxwell-Jia/Spec-o3-ColdStartSFT.swesmith-coldstart-rl-seqnorm-tis-pym2tmc_gui_cold_startMetis-ColdStart
Metis-ColdStart
Act Wisely: Cultivating Meta-Cognitive Tool Use in Agentic Multimodal Models
Metis-ColdStart is the supervised fine-tuning (SFT) dataset used to train the Metis-8B-ColdStart model. It contains ~27K high-quality, tool-augmented multimodal reasoning trajectories that have been rigorously curated to ensure genuine tool necessity and reasoning quality.
[Paper (arXiv)] | [GitHub] | [ColdStart Model] | [RL Model] | [RL Data]
Dataset Details
Attribute
Value… See the full description on the dataset page: https://huggingface.co/datasets/Accio-Lab/Metis-ColdStart.SecCoderX_Reasoning_Vulnerability_Detection_SFT_Cold_Start_Dataset
Citation
If you find our work helpful, feel free to give us a cite.
@misc{wu2026securecodegenerationonline,
title={Secure Code Generation via Online Reinforcement Learning with Vulnerability Reward Model},
author={Tianyi Wu and Mingzhe Du and Yue Liu and Chengran Yang and Terry Yue Zhuo and Jiaheng Zhang and See-Kiong Ng},
year={2026},
eprint={2602.07422},
archivePrefix={arXiv},
primaryClass={cs.CR},
url={https://arxiv.org/abs/2602.07422}… See the full description on the dataset page: https://huggingface.co/datasets/SecCoderX/SecCoderX_Reasoning_Vulnerability_Detection_SFT_Cold_Start_Dataset.Mini-o3-Coldstart-DatasetOpenThoughts-Agent-SFT-ColdStartForRL-10K
Project |
Code |
Collection
OpenThoughts-Agent-SFT-ColdStartForRL-10K
OpenThoughts-Agent is an open-source effort to curate the best datasets for training agents. Our release includes datasets, models and our research codebase.
OpenThoughts-Agent-SFT-ColdStartForRL-10K is the cold-start supervised-finetuning dataset for the OpenThoughts-Agent SFT→RL recipe. It contains 9,437 (task, agent-trajectory) pairs that teach a base model the agentic format and behaviour… See the full description on the dataset page: https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-SFT-ColdStartForRL-10K.SciLM-CoT_ColdStart
SciReasoner: Laying the Scientific Reasoning Ground Across Disciplines
This repo contains the cold-start data of SciReasoner.
drkernel-coldstart-8k
DR.Kernel Cold-Start Dataset
Paper | Code
This directory documents the format of hkust-nlp/drkernel-coldstart-8k.
The cold-start set is used for supervised fine-tuning (SFT) before RL in DR.Kernel. As described in the paper, it is built from 5-turn multi-turn trajectories collected with KernelGYM feedback.
Overview
Purpose: initialize kernel-generation ability (Triton coding + iterative optimization) before TRLOO/MRS/PR/PRS RL.Data form: one row per full multi-turn… See the full description on the dataset page: https://huggingface.co/datasets/hkust-nlp/drkernel-coldstart-8k.OVR-Dataset-ColdStarttriageiq-dataset
TriageIQ — Customer Support Ticket Classification Dataset
Synthetic dataset of 2,000 customer support tickets, each labeled with three independent classification axes.
Schema
Each example is a JSON object:
{
"text": "I've been charged twice this month, please refund me ASAP.",
"sentiment": "negative", // positive | neutral | negative
"urgency": "high", // low | medium | high
"category": "billing" // billing | technical |… See the full description on the dataset page: https://huggingface.co/datasets/coldstart88/triageiq-dataset.Multimodal-Cold-Start
Multimodal-Cold-Start Dataset: Advancing Multimodal Reasoning via Reinforcement Learning with Cold Start
This repository hosts the Multimodal-Cold-Start dataset, a crucial component of the research presented in the paper "Advancing Multimodal Reasoning via Reinforcement Learning with Cold Start". This dataset is specifically designed for the cold start supervised fine-tuning (SFT) stage, which serves as the initial phase in a two-stage approach to enhance multimodal reasoning… See the full description on the dataset page: https://huggingface.co/datasets/WaltonFuture/Multimodal-Cold-Start.GrepSeek-ColdStart-SFT-10k
GrepSeek — Cold-start SFT dataset (10k)
Cold-start supervised fine-tuning data for GrepSeek, a Direct Corpus
Interaction (DCI) search agent. Instead of retrieving from a pre-computed dense
or sparse index, a DCI agent answers questions by issuing Unix shell commands
(rg, grep, head, …) directly against a raw text corpus — coupling retrieval
and reasoning in a single policy. These trajectories are used to initialize the
agent before reinforcement learning.
📄 Paper: GrepSeek:… See the full description on the dataset page: https://huggingface.co/datasets/alireza7/GrepSeek-ColdStart-SFT-10k.Cos-Play-Cold-Start
COS-PLAY Cold-Start Data
Pre-generated cold-start data for COS-PLAY (COLM 2026): Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Game Play.
📄 Paper: arXiv:2604.20987 · HuggingFace Paper Page
💻 Code: github.com/wuxiyang1996/cos-play
🌐 Project page: wuxiyang1996.github.io/COSPLAY_page
🤖 Models: IntelligenceLab/COS-PLAY
Dataset Summary
This dataset contains GPT-5.4-generated seed trajectories and skill-labeled episodes for 8 games, used to bootstrap… See the full description on the dataset page: https://huggingface.co/datasets/IntelligenceLab/Cos-Play-Cold-Start.AVAR-ColdStart-30KGeoVista-Cold-Starttb2-sft-coldstart-decontam
tb2-sft-coldstart-decontam
Decontaminated Terminus-2 SFT cold-start trajectories from nvidia/Nemotron-Terminal-Corpus (DeepSeek-V3.2), with every held-out Terminal-Bench task removed by n-gram phrase containment + the TB canary GUID. messages is a JSON chat array (assistant turns = loss-masked SFT targets). See decontam_report.json for the contamination rate per held-out task.
llm-cold-start-benchmark
LLM Container Cold-Start Benchmark
Measurements of how long it takes to bring a language model from cold storage to
a state where it can serve its first token, across 25 open-weight
models spanning 17 architecture families and
100.9 GiB of checkpoints, on a single NVIDIA T4.
Cold start is the latency a serverless or scale-to-zero inference platform pays
when it has no warm replica. It decomposes into weight transfer from storage,
deserialization into host memory, transfer to the… See the full description on the dataset page: https://huggingface.co/datasets/ArchCoder/llm-cold-start-benchmark.GThinker-11k-cold-start
Introduction
This repository contains the cold-start dataset used in the paper GThinker: Towards General Multimodal Reasoning via Cue-Guided Rethinking.
We build this data using our proposed Multimodal Iterative Labeling framework, with the help of O1, O3, and GPT-4o. To use this dataset, you must follow the policy of OpenAI and M3CoT and cite as follows.
Guideline
TRAIN_SYS_PROMPT = 'A conversation between User and Assistant. The user asks a question, and the Assistant… See the full description on the dataset page: https://huggingface.co/datasets/JefferyZhan/GThinker-11k-cold-start.minio3-qa-coldstart-2606eval-laion_swesmith-coldstart-complete-lt32k-2ep-8B_DCAgent_dev_set_v2-tracesmath-cold-startterminal_bench_2_swesmith_coldstart_rl_seqnorm_tis_pym2t_75_8B_20260703_054925physr1corp-cold-start
PhysR1Corp Cold-Start — Tool-Using Physics Trajectories (Text + Multimodal)
SFT cold-start dataset for Physics-R2 / Phase E (a tool-use RL paper for VLMs, target ICLR 2027).
1,973 audited trajectories — 1,740 text-only + 233 multimodal — covering the full PhysR1Corp
corpus (all 2,268 problems). Each trajectory solves a physics problem using a custom SymPy tool
routed through the project's harness.sandbox runtime. The trajectory schema matches the W6
RL-training inference format… See the full description on the dataset page: https://huggingface.co/datasets/shanyangmie/physr1corp-cold-start.R3-RAG-ColdStartTrainingDataswesmith-coldstart-2ep-8B-swebench-traceswalton-multimodal-cold-start-r1-format-30k
walton-multimodal-cold-start-r1-format-30k
WaltonFuture/Multimodal-Cold-Start converted to multimodal-open-r1-8k-verified format with filtering
Dataset Description
This dataset was processed using the data-preproc package for vision-language model training.
Processing Configuration
Base Model: Qwen/Qwen2.5-7B-Instruct
Tokenizer: Qwen/Qwen2.5-7B-Instruct
Sequence Length: 16384
Processing Type: Vision Language (VL)
Dataset Features
input_ids:… See the full description on the dataset page: https://huggingface.co/datasets/penfever/walton-multimodal-cold-start-r1-format-30k.terminal_bench_2_swesmith_coldstart_rl_seqnorm_tis_muon_pym2t_25_8B_20260703_054927search-cold-start