shreethar/FYP-Stage2-dataset

🤖 FYP Stage 2 — VLA Pre-training Subset

A compact, self-contained multi-source dataset for Vision-Language-Action (VLA) Stage 2 pre-training. Built as a portable ~90 GB subset of 8 larger upstream sources, it is designed to be pulled via a single load_dataset() call — no raw 100+ GB downloads required at training time.

All images and video frames are pre-materialized at 448 × 448 resolution and stored inline.

---

📊 Dataset at a Glance

Split	Samples	Size
train	44,170	~78 GB
test	7,102	~13 GB
Total	51,272	~90 GB

Split ratio: 85 % train / 15 % test (deterministic — same record always lands in the same split).

---

🗂️ Data Sources

Source	Task Type	Media	Target Count
MolmoAct	Trajectory prediction	Single image	10 000
ShareRobot Affordance	Affordance bbox	Single image	~6 500 (all)
ShareRobot Planning	Task planning QA	Multi-frame	10 000
RoboVQA	Robot VQA	Multi-frame video	10 000
Pixmo Cap	Image captioning	Single image	2 000
Pixmo AMA	Open-ended QA	Single image	2 000
Pixmo Cap-QA	Caption-grounded QA	Single image	2 000
RoboFAC	Failure analysis QA	Multi-frame video	10 000

---

🔍 Schema

Column	Type	Description
dataset	string	Source identifier (e.g. "molmoact", "robofac")
type	string	Task category — "trajectory" or "qa"
human	string	User turn prompt (system prompt + task instruction)
assistant	string	Ground-truth response
frames	List[Image]	1 frame (image tasks) or up to 16 frames (video tasks), 448×448 RGB
media_type	string	"image" (1 frame) or "video" (> 1 frame)

---

🚀 Usage

from datasets import load_dataset

ds = load_dataset("shreethar/FYP-Stage2-dataset")
train_ds = ds["train"]
test_ds  = ds["test"]

# Inspect a sample
sample = train_ds[0]
print(sample["dataset"])    # e.g. "molmoact"
print(sample["media_type"]) # "image" or "video"
print(sample["human"])      # user prompt
print(sample["assistant"])  # ground-truth answer
sample["frames"][0].show()  # PIL Image

Filter by source

robofac = train_ds.filter(lambda x: x["dataset"] == "robofac")

Filter by media type

video_samples = train_ds.filter(lambda x: x["media_type"] == "video")
image_samples = train_ds.filter(lambda x: x["media_type"] == "image")

Reconstruct chat messages (Qwen-VL format)

def to_messages(sample):
    media = sample["frames"]
    if sample["media_type"] == "video":
        media_block = {"type": "video", "video": media}
    else:
        media_block = {"type": "image", "image": media[0]}
    return [
        {"role": "user",      "content": [media_block, {"type": "text", "text": sample["human"]}]},
        {"role": "assistant", "content": sample["assistant"]},
    ]

---

🏗️ Construction

This dataset was built by the build_hf_subset.py script in the training repository. The pipeline runs in four phases:

1. Metadata indexing — lightweight record dicts (paths, offsets, text) per source
2. Materialization — images/video frames are loaded, decoded, and resized to 448 × 448
3. Splitting — deterministic 85/15 split via MD5 hash of (source, prompt, response)
4. Push — DatasetDict.push_to_hub()

Deterministic sampling (_keep) and splitting (_get_split) ensure identical records are selected and assigned to the same split on every rebuild.