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 rule reward and a concise reference configuration for the RL experiment.

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

Judge: SUCCESS
Judge: FAIL

Layout

sft/datasets/                 SFT JSON files
sft/viewer/                   Parquet view of the SFT data for Hub preview
sft/images-shards/            independent image tar shards
sft/dataset_info.json         LLaMAFactory dataset entries
rl/train.parquet              GRPO training split
rl/val.parquet                GRPO validation split
rl/osreward_reward.py         rule reward
stats/                        compact release statistics

Quick Start

Extract all screenshots from the repository root:

bash sft/extract_images.sh

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:

sha256sum -c SHA256SUMS

No model weights or training framework checkout is included.