datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
OpenHand-Synth
Dataset Card for OpenHand-Synth
📜 Paper: OpenHand-Synth: A Large-Scale Synthetic Handwriting Dataset for Multimodal Language Models
Sample Images
Image
Ground Truth
Source
Language
CER
JW
02-10-1436
faker-date
por
0.10
0.96
Stephan Thomsen-Johansen
faker-name
dan
0.0
1.0
Le chat mange.
tatoeba
fra
0.0
1.0
Classical musicsoothes me.She took the risk, knowing that shemight lose a lot of money.I could not catcha single word of their talk.In the old days… See the full description on the dataset page: https://huggingface.co/datasets/to-be/OpenHand-Synth.openhands-feedback
OpenHands Feedback Dataset 🙌
Dataset Description
What is OpenHands Feedback?
The OpenHands Feedback Dataset is a collection of user interactions and feedback with the OpenHands AI coding assistant. This dataset contains real-world examples of how users interact with AI coding assistants, including both successful and unsuccessful interactions, along with user feedback on the quality and helpfulness of the responses.
The dataset currently contains 275 examples… See the full description on the dataset page: https://huggingface.co/datasets/OpenHands/openhands-feedback.openhands-divergence-dpo-strong
openhands-divergence-dpo-strong
Strong-preference subset of divergence-point DPO pairs for OpenHands-style
tool use. Filtered for decisive exclusivity and chosen-strong actions
(edit/create/test/make_test/script; search only when the rejected side is
explore). Soft explore↔explore junk is out.
Current revision: strong-v2.2 (local soft post-filter after Round 12
spot-check fail). Remine quality bars unchanged from v1/v2. Volume from 8+8
re-spill; quality recovery drops… See the full description on the dataset page: https://huggingface.co/datasets/asaverren/openhands-divergence-dpo-strong.tool-reasoning-sft-RESEARCH-OpenHands-CodeScout_Training_Rollouts
CodeScout Training Rollouts — Cleaned & Rectified
~40K multi-turn code localization agent trajectories converted into a strict reasoning + tool-call format with validated FSM transitions. Supports coupled (parallel) tool calls.
⚠️ Mid-training dataset. This dataset contains synthesized reasoning templates (not native chain-of-thought). It is suitable for mid-training to teach tool-use mechanics, FSM structure, and bash exploration patterns. It is not recommended as a final SFT… See the full description on the dataset page: https://huggingface.co/datasets/AmanPriyanshu/tool-reasoning-sft-RESEARCH-OpenHands-CodeScout_Training_Rollouts.swerebench-openhands-100m-max64k
SWE-rebench OpenHands 100M SFT Subset (max 64k)
This is a deterministic, representative subset of
nebius/SWE-rebench-openhands-trajectories,
augmented with exact sequence and supervised-loss token counts. The source trajectories were
collected with Qwen3-Coder-480B-A35B-Instruct and OpenHands v0.54.0. This derivative preserves
the source dataset's CC BY 4.0 license and attribution.
Filters and size
7,867 trajectories
100,095,655 assistant loss tokens
352,709,237… See the full description on the dataset page: https://huggingface.co/datasets/hanspeterlyngsoeraaschoujensen/swerebench-openhands-100m-max64k.openhands-train-ready
openhands-train-ready
SFT-ready normalization of
nebius/SWE-rebench-openhands-trajectories
— 5,000 resolved OpenHands (v0.54.0) agent trajectories on real GitHub issues,
bootstrapped by Qwen/Qwen3-Coder-480B-A35B-Instruct, with tool-call
arguments deserialized to objects so chat templates just work.
What was fixed
The upstream parquet stores tool_call.function.arguments as a serialized
JSON string "for storage efficiency" (per the Nebius README). Chat templates… See the full description on the dataset page: https://huggingface.co/datasets/asaverren/openhands-train-ready.openhands-divergence-dpo
openhands-divergence-dpo
5,420 divergence-point DPO pairs mined from
nebius/SWE-rebench-openhands-trajectories
(67,074 OpenHands v0.54.0 trajectories by Qwen3-Coder-480B-A35B-Instruct on real
GitHub issues from SWE-rebench).
Where Nebius released raw trajectories for RFT/RL, this dataset extracts
preference pairs at the first diverging action: for GitHub issues attempted
multiple times where at least one attempt resolved the issue and at least one
failed, we align a resolved and… See the full description on the dataset page: https://huggingface.co/datasets/asaverren/openhands-divergence-dpo.filtered-SWE-Openhands-Devstral-32k-20K
Dataset Summary
This dataset contains 20,000 high-quality SFT trajectories filtered and selected from SWE-Star, a 250K+ sample dataset collected under a modified OpenHands scaffold. The data is used to train SWE agent models under limited resources with a 32K context window.
A single-source dataset is intentionally selected so that all training trajectories share a uniform scaffold, isolating the scaffold inconsistency effect: any performance variation across evaluation… See the full description on the dataset page: https://huggingface.co/datasets/ubicloud/filtered-SWE-Openhands-Devstral-32k-20K.ClaudeCode-OpenHands
ClaudeCode in OpenHands Native trajectory format
About AgentIR Collection
This dataset is part of the AgentIR Collection.
AgentIR is an open-source compiler infrastructure for agentic trajectories (like LLVM/MLIR, but for agent traces).
Using AgentIR, you can convert any source trajectory format into multiple target formats.
Project: https://github.com/ravenSanstete/agentir
DSL: Define custom formats with *.agentir.yaml files
CLI: agentir dsl convert for… See the full description on the dataset page: https://huggingface.co/datasets/WhitzardAgent/ClaudeCode-OpenHands.AgentTrove-OpenHands
AgentTrove in OpenHands Native trajectory format
About AgentIR Collection
This dataset is part of the AgentIR Collection.
AgentIR is an open-source compiler infrastructure for agentic trajectories (like LLVM/MLIR, but for agent traces).
Using AgentIR, you can convert any source trajectory format into multiple target formats.
Project: https://github.com/ravenSanstete/agentir
DSL: Define custom formats with *.agentir.yaml files
CLI: agentir dsl convert for one-command… See the full description on the dataset page: https://huggingface.co/datasets/WhitzardAgent/AgentTrove-OpenHands.