#!/usr/bin/env python3
"""Build preview.html: embedded sample videos per regime + their questions + stats."""
import base64
import csv
import json
from collections import Counter
from pathlib import Path
ROOT = Path(__file__).parent / "dataset"
REGIME_TITLES = {
"train": "TRAIN (in-distribution, matches pendulum-cjepa-160f recipe)",
"val_high_damp": "VAL OOD · high damping γ∈(0.55,0.90)",
"val_low_damp": "VAL OOD · near-zero damping γ∈(0,0.08)",
"val_large_angle": "VAL OOD · large release angle ±(45–80°)",
"val_gravity": "VAL OOD · Moon / Mars / Jupiter gravity",
"val_quad_damping": "VAL OOD · quadratic drag β·ω|ω| (new law)",
"val_driven": "VAL OOD · periodically driven pendulum (new forcing)",
}
def b64(p):
return base64.b64encode(open(p, "rb").read()).decode()
def esc(s):
return str(s).replace("&", "&").replace("<", "<").replace(">", ">")
def main():
idx = list(csv.DictReader(open(ROOT / "index.csv")))
by_regime = {}
for r in idx:
by_regime.setdefault(r["regime"], []).append(r)
qj = {}
for split in ("train", "val"):
for e in json.load(open(ROOT / "questions" / f"questions_{split}.json"))["scene_annotation"]:
qj[e["video_id"]] = e["questions"]
stats_type = Counter()
stats_concept = Counter()
for qs in qj.values():
for q in qs:
stats_type[q["question_type"]] += 1
stats_concept[q.get("concept", "?")] += 1
parts = ["""
Pendulum 3K · CLEVRER-style QA dataset — preview
3,000 videos (160 frames @12 fps, 196×196, H.264) · frame-aligned CSV labels ·
~24k questions across the four CLEVRER types. Validation is out-of-distribution
physics on purpose (causal generalization), not just a held-out seed.
All question answers are derived from the ground-truth RK4 simulation
(videos show 13.33 s; a hidden 28 s continuation grounds the predictive questions).
"""]
n_q = sum(len(v) for v in qj.values())
parts.append(f"Dataset stats
| videos | questions |
|---|
"
f"| train (in-distribution) | 2400 | — |
"
f"| val (6 OOD regimes × 100) | 600 | — |
"
f"| total | 3000 | {n_q} |
")
parts.append("| question type | count |
|---|
"
+ "".join(f"| {k} | {v} |
" for k, v in stats_type.most_common())
+ "
")
parts.append("causal-concept histogram
"
"| concept | count |
|---|
"
+ "".join(f"| {esc(k)} | {v} |
" for k, v in stats_concept.most_common())
+ "
{esc(REGIME_TITLES[regime])}
")
params = (f"θ₀={float(r['theta0_deg']):+.1f}° ω₀={float(r['omega0_rad_s']):+.3f} "
f"γ={float(r['damping_s_inv']):.3f} β={float(r['quad_drag_beta']):.3f} "
f"g={float(r['gravity_m_s2']):.2f} drive_A={float(r['drive_amp_rad_s2']):.2f} "
f"f_d={float(r['drive_freq_hz']):.3f}Hz")
parts.append(f"{vid}.mp4 · {esc(params)}
")
parts.append('')
parts.append(f'
')
parts.append('
')
for q in qj[vid]:
parts.append(f'- {q["question_type"]}{esc(q["question"])}
')
if "choices" in q:
for i, c in enumerate(q["choices"]):
cls = ' class="c"' if i == q["answer"] else ""
parts.append(f'{"abcd"[i]}) {esc(c)} · ')
else:
parts.append(f'answer: {esc(q["answer"])}')
parts.append(" ")
parts.append("