Physical-ICL

Physical-ICL is a project-level dataset repository for studying physical in-context learning in video generation.

The current version contains a preliminary subset built from Physics-IQ. Each sample is organized as a query video with candidate demonstration videos. These demonstrations are labeled by their relationship to the query, such as good, weak, opposite, or irrelevant demonstrations.

All 345 task records and the first ordinary SFT (including final LoRA and 32 historical videos) are now included: All 345 tasks · First SFT.

H3 causal-ICL release — 2026-10-17

data/h3_causal_icl_v1/ contains the current 5,341-pair causal rulebank, the completed outcome-hidden diagnostic (65 train / 2 dev, two independent 200-update LoRA models, 12 videos), exact model inputs, losses, quality flags and a browsable code mirror. Rubrics and dev evidence are separate from model inputs. Eight mechanism-positive training pairs have different specific responses and are explicitly flagged. This two-case diagnostic does not establish an overall causal-ICL gain.

The explicit h3_outcome_hidden_primary_v1 dataset-viewer configuration loads only its 65/2 task manifests; the older media collections remain available under their existing paths. Original training source videos are referenced, not duplicated. Browse code · 12 generated videos · Review notes.

Current subset

Subset	Path	Description
Physics-IQ preliminary subset	data/physiq_prelim/	A preliminary physical ICL dataset constructed from Physics-IQ videos.

Repository structure

data/
  physiq_prelim/
    gt_data/
      task_0001/
        episode_0001/
          video.mp4
          prompt/
            init_frame.png
            prompt.txt
          demos/
            good_demo_01.mp4
            good_demo_01.png
            weak_demo_01.mp4
            weak_demo_01.png
            opposite_demo_01.mp4
            opposite_demo_01.png
            irrelevant_demo_01.mp4
            irrelevant_demo_01.png
    summary.json
    case_summary.csv
    README.md
    export_warnings.txt
    storyboard_warnings.txt

Sample format

Each sample is stored under:

data/physiq_prelim/gt_data/task_xxxx/episode_0001/

The files have the following meanings:

File or folder	Description
video.mp4	Query target video. In the current version, this uses the full Physics-IQ video when available.
prompt/init_frame.png	Query initial frame, extracted from the first frame of the full query video.
prompt/prompt.txt	Text prompt for the query video.
demos/*.mp4	Candidate demonstration videos. These use 5-second Physics-IQ testing clips.
demos/*.png	3x3 event-aware storyboard images generated from the corresponding demo video.

Demo types

Demo files are named by their coarse relationship to the query:

Filename pattern	Meaning
good_demo_XX.mp4	A suitable positive demonstration.
weak_demo_XX.mp4	A weakly related demonstration.
opposite_demo_XX.mp4	A demonstration showing an opposite or contrastive physical outcome.
irrelevant_demo_XX.mp4	An unrelated or different-category control demonstration.

The corresponding .png file is a 3x3 storyboard extracted from the same demo video. For example:

good_demo_01.mp4
good_demo_01.png

Low-quality generated demonstrations are not included in the current version.

Metadata files

File	Description
data/physiq_prelim/summary.json	Machine-readable metadata for all samples and demonstrations.
data/physiq_prelim/case_summary.csv	Human-readable case-level summary.
data/physiq_prelim/export_warnings.txt	Export warnings, if any.
data/physiq_prelim/storyboard_warnings.txt	Storyboard generation warnings, if any.

Metadata schema

Each item in summary.json corresponds to one query sample. The main fields are:

Field	Description
case_id	Unique case identifier.
task_name	Task folder name.
gt_path	Path to the query video.
image	Path to the query initial frame.
prompt	Query prompt.
query_scenario	Physics-IQ scenario name for the query.
query_macro_group	Coarse physical category.
query_event_tag	Fine-grained event tag.
demos	Candidate demonstrations for this query.
available_demo_types	Available demo types for this query.

Each demo entry contains:

Field	Description
demo_type	One of good, weak, opposite, or irrelevant.
demo_path	Path to the demo video.
demo_image_path	Path to the 3x3 storyboard image.
demo_scenario	Physics-IQ scenario name for the demo.
demo_relation	More detailed relation label.
physical_similarity	Physical similarity label.
visual_similarity	Visual similarity label.

Usage example

Load the metadata:

import json
from pathlib import Path

root = Path("data/physiq_prelim")

with open(root / "summary.json", "r", encoding="utf-8") as f:
    items = json.load(f)

sample = items[0]

query_video = root.parent.parent / sample["gt_path"]
query_image = root.parent.parent / sample["image"]
query_prompt = sample["prompt"][0]

good_demos = [
    d for d in sample["demos"]
    if d["demo_type"] == "good"
]

For video-capable models, use demo_path. For image-only models, use demo_image_path.

Notes

• Query videos use full Physics-IQ videos when available.
• Demo videos use 5-second Physics-IQ testing clips.
• Demo storyboard images are generated using event-aware 3x3 frame sampling.
• The current version does not include low-quality generated demonstrations.
• This repository is intended for research and preliminary experiments.