liangyuch/ttcc-v0_2_0

TikTok Creative Center — Top Ads (v0.2.0, full corpus)

Full corpus scraped from the TikTok Creative Center Top Ads surface (country=US, period=180, sort=for_you). One row per ad_id with embedded video bytes (mp4), cover image (jpg/png), per-second engagement curves, and TikTok-reported metadata.

61,789 ads. Sister dataset liangyuch/ttcc-v0_1_0 is the legacy 935-ad smoke set with the same schema — use it for quick sanity-checks before pulling the full v0.2.0.

Splits

Deterministic 80 / 10 / 10 via SHA256("v0.2.0:" + ad_id) % 100. Splits are stable: re-exports against the same version_seed land each ad in the same bucket forever.

• train: 49,479 ads
• val: 6,114 ads
• test: 6,196 ads

Load with:

from datasets import load_dataset
train = load_dataset("liangyuch/ttcc-v0_2_0", split="train")
val   = load_dataset("liangyuch/ttcc-v0_2_0", split="val")
test  = load_dataset("liangyuch/ttcc-v0_2_0", split="test")

Schema (51 columns)

One row per ad_id. Columns mirror v0.1.0 exactly.

• identity: ad_id, vid, ad_title, brand_name, industry_key (+ industry_name), objective_key (+ objective_name), source, source_key, list_country_code
• assets (embedded bytes): video_local_path (HF Video), cover_local_path (HF Image, may be NULL for ads where cover failed to download)
• detail metadata: caption, objectives_detail, objectives_named, keyword_list, pattern_label, landing_page, countries_delivered, voice_over, highlight_text, has_summary, is_search, favorite, detail_page_url
• video properties: duration, width, height
• engagement: ctr_global (percentile rank, lower=better), percentile_global, cost_bucket (0/1/2), like_cnt, comments_cnt, shares_cnt, metrics_period
• per-second curves (1 Hz): ctr_curve, cvr_curve, clicks_curve, conversion_curve, retention_curve
• provenance: first_seen_at, last_seen_at, video_downloaded_at, cover_downloaded_at, detail_scraped_at, video_error, cover_error, detail_error, video_refresh_attempts, split

Filter applied (v2.6 relaxed)

detail_scraped_at IS NOT NULL
AND (detail_error IS NULL OR detail_error LIKE 'partial_curves:%')
AND ctr_curve       IS NOT NULL
AND retention_curve IS NOT NULL
AND video_local_path IS NOT NULL

Keeps partial_curves:N/5 rows (ctr + retention present); tolerates missing cover/caption.

Caveats

• ctr_global is a percentile rank, not a click-through rate. Lower is better.
• cost_bucket is an ordinal (0/1/2). ~85% are High in top-ads.
• Curves are per playback second, not calendar time.
• ad_title (short headline) ≠ caption (full on-screen copy).
• objective_key reflects the slice that surfaced the ad; the ad's full objective set is in objectives_detail (often multi-valued).
• Some ad_ids may have been delisted by TikTok after the crawl; those rows have populated curves but caption / objectives_detail may be NULL.
• partial_curves:N/5 rows have 2-4 of 5 curves populated (ctr + retention always present by filter; the other 3 may be NULL).
• Videos can be partially watermarked (Spark Ads).

Provenance

• Source: ads.tiktok.com/business/creativecenter (publicly browsable).
• Scraped via Playwright route-interception, v2.6 relaxed filter.
• Window: 2026-04 → 2026-05.
• Research-use only. Per TikTok's general Terms of Service, automated collection of platform data is restricted; this set is shared as an academic artifact for creative-engagement research. Not for commercial use, redistribution, or production model training.