EgoAction-SFT-5K contains 5,354 egocentric video episodes paired with English instructions and temporally grounded, hand-specific action captions for supervised fine-tuning. Each response describes the camera wearer's left-hand, right-hand, and coordinated two-hand actions.

This release preserves the SFT v1 training data used by the paper's EgoAction (LLaVA-OV2-8B) model. Its training receipt identifies the same frozen data prepared for the Qwen3-VL-4B SFT v1 run. The OV2 model used checkpoint-322 from that training run. This release is tied to those verified training files; it does not assert a checkpoint mapping for other paper rows.

Dataset composition

Split	Episodes
Train	5,146
Validation	208
Total	5,354

Video source	Episodes
Ego4D	2,145
EPIC-KITCHENS	1,932
EgoLife	1,277

The original split assignments are preserved. Recording groups recorded in the source manifests are disjoint between train and validation. The videos occupy 51,027,915,845 bytes (approximately 51.0 GB) and are included without re-encoding.

Annotation and prompt format

The SFT v1 assistant response is a JSON string with three arrays: left_hand, right_hand, and both_hands. Each segment contains start_s, end_s, and caption. Times are in seconds relative to the video clip. Captions are English action phrases. This historical training format includes placeholder segments such as idle, not visible, and no two-hand collaboration; these are preserved in the released targets.

The exact training prompt is in provenance/v1/prompt.txt. The instruction and assistant response of every episode retain their original text. The publication changes video references to repository-relative paths and adds a tabular convenience view. It does not change target text, timestamps, track assignments, video bytes, or train/validation membership.

These are model-assisted training labels, including translated English targets. They are not the independently human-annotated benchmark ground truth. Source label status, translation provenance, and original hashes remain available in the provenance manifest. Publication does not confer additional annotation QA.

Files

Path	Content
`data/v1/{train,validation}.jsonl`	Convenient rows with episode ID, video path, source, duration, recording group, prompt, and response
`formats/v1/{train,validation}.trainer.jsonl`	Actual training rows with portable video paths and original `chat_template_kwargs`
`formats/v1/{train,validation}.messages.jsonl`	Typed multimodal user messages and assistant responses
`formats/v1/{train,validation}.sharegpt.jsonl`	Video-placeholder chat messages with a `videos` list
`videos/*.mp4`	All 5,354 original video clips, shared by all views
`provenance/v1/`	Original training files and manifest, training prompt, audit, and OV2 source receipt
`dataset_statistics.json`	Split, source, and segment counts
`release_manifest.json`	File sizes and SHA-256 checksums for the release payload
`prepare_training.py`	Materialize trainer JSONL with absolute paths after download

The training-file hashes recorded by the OV2 run are:

Train: 628e66939d7e5c303536fbf769005e82c0ebf45b2ed146c3a979b4032a217ad3
Validation: 980d8944711dd2bbcc947c63c76059e5fd4127df3411c41f3238aa1a101979f8

Decompressing provenance/v1/*.trainer.original.jsonl.gz reproduces those exact files. Historical absolute paths in provenance describe the original training environment; use data/, formats/, and the helper for portable use.

Access and download

This repository is public with manual access approval. Request access on this page, wait for approval, and authenticate with your approved Hugging Face account. The dataset owner reviews requests in the repository settings.

Download all annotations and videos:

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="play0718/EgoAction-SFT-5K",
    repo_type="dataset",
    token=True,
    local_dir="EgoAction-SFT-5K",
)

Prepare the training and validation JSONL with local absolute video paths:

python EgoAction-SFT-5K/prepare_training.py --output prepared-egoaction-sft

The resulting train.jsonl and validation.jsonl retain the original training messages, targets, and video-processing parameters. Choose the model's own chat template and video processor, and apply loss to the assistant response. The original trainer rows record 10 fps, max_frames=512, min_pixels=65536, max_pixels=131072, and total_pixels=33554432; model-specific training code determines how these settings are consumed.

For annotation-only exploration with Hugging Face Datasets:

from datasets import load_dataset

dataset = load_dataset("play0718/EgoAction-SFT-5K", "v1", token=True)
example = dataset["train"][0]
# example["video"] is a repository-relative path, not downloaded video bytes.

Source terms

Source videos originate from Ego4D, EPIC-KITCHENS, and EgoLife. Their respective source terms continue to apply; repository access approval does not replace those terms or grant additional rights to the underlying videos.

Release maintained by play0718.

Downloads last month: 788

Total file size:

200 GB