Datasetpublic

OS-Copilot/OS-Shepherd-100K

OSReward Data OSReward is a multimodal reward-model dataset for judging whether GUI-agent trajectories complete the user's task. The release contains: Supervised fine-tuning data in a LLaMAFactory-compatible ShareGPT layout. Deduplicated screenshots packaged in independently extractable tar shards. Training and validation data for GRPO. The expected answer contains a brief evidence-based analysis followed by a final line in exactly one of these forms: Judge: SUCCESS Judge:… See the full description on the dataset page: https://huggingface.co/datasets/OS-Copilot/OS-Shepherd-100K.

sourceHugging Faceupdated 26d agoView on Hugging Face

4likes452downloads

Dataset Card

OSReward Data

OSReward is a multimodal reward-model dataset for judging whether GUI-agent trajectories complete the user's task. The release contains:

—Supervised fine-tuning data in a LLaMAFactory-compatible ShareGPT layout.
—Deduplicated screenshots packaged in independently extractable tar shards.
—Training and validation data for GRPO.

The expected answer contains a brief evidence-based analysis followed by a final line in exactly one of these forms:

text

Judge: SUCCESS
Judge: FAIL

Layout

text

sft/datasets/                 SFT JSON files
sft/viewer/                   Parquet view of the SFT data for Hub preview
sft/images-shards/            independent image tar shards
rl/train.parquet              GRPO training split
rl/val.parquet                GRPO validation split
stats/                        compact release statistics

Quick Start

Extract all screenshot shards from the repository root:

bash

for shard in sft/images-shards/*.tar; do tar -xf "$shard"; done

This restores paths under osreward_rm_train_bundle/images/, matching the relative image paths used by both SFT and RL records.

The SFT JSON files are the canonical LLaMAFactory training files. The Parquet files under sft/viewer/ contain the same records in a format supported by the Hub dataset preview.

See DATA_CARD.md, sft/SFT_FORMAT.md, and rl/RL_FORMAT.md for schemas and limitations.

Integrity

Verify the downloaded release with:

bash

sha256sum -c SHA256SUMS

No model weights or training framework checkout is included.

WebTrail 100K completion increment

The release now contains exactly 100,000 SFT judge instances. The additive WebTrail subset contributes 6,386 unique web trajectories using the last up to five screenshots (last5); trajectories shorter than five states naturally contain fewer images. Its canonical files are sft/datasets/osreward_webtrail_supplement_single_filter.json and sft/viewer/osreward_webtrail_supplement_single_filter.parquet.

The 32 independently extractable WebTrail image shards are described by sft/webtrail_image_shards_manifest.json. The existing extraction loop over sft/images-shards/*.tar includes both the original and WebTrail shards.