agerating

Suggested minimum ages for about 600,000 films and TV shows on Wikidata, published every week as SQLite databases. About 20,000 of the ages are official FSK ratings from Wikidata; the rest are estimates from a fine-tuned language model.

Ages are on the German FSK scale: 0, 6, 12, 16, 18. Estimates are machine predictions, not ratings. Never present them as FSK or other official ratings; the license requires this.

Downloads

Each release has two databases, compressed with zstd, the small one's lookups as a gzipped packed list, and their checksums in SHA256SUMS:

File	Contents	Size
agerating-ids.db.zst	small: IMDb, TMDB and TheTVDB ID → age	2.3 MB (10 MB unpacked)
agerating-ids.bin.gz	the small database's lookups as a packed list, gzip	0.8 MB (2 MB unpacked)
agerating.db.zst	full: every title with its external IDs, official rating, estimate and model probabilities	164 MB (567 MB unpacked)

gh release download --repo edde746/agerating --pattern agerating-ids.db.zst
zstd -d agerating-ids.db.zst
sqlite3 agerating-ids.db "SELECT age FROM imdb WHERE id = 133093"   # The Matrix (tt0133093) -> 16

The small database

One table per ID type, each (id INTEGER PRIMARY KEY, age INTEGER NOT NULL):

Table	ID	Example
imdb	IMDb title ID as a number, without tt and leading zeros	tt0133093 → 133093
tmdb_movie	TMDB movie ID	603
tmdb_tv	TMDB TV series ID	1399
tvdb_series	TheTVDB series ID	121361
tvdb_movie	TheTVDB movie ID

TMDB and TheTVDB number movies and series separately, so look IDs up in the matching table. age is the official FSK rating where Wikidata has one, otherwise the estimate. When an ID belongs to several Wikidata items, an official rating wins, otherwise the highest estimate. IDs that are not plain numbers on Wikidata (URLs, slugs) are left out.

The packed list

agerating-ids.bin.gz holds the same IDs and ages as the small database in a form an app can read with nothing but a gzip decoder. Integers are little-endian; a varint is LEB128 (7 bits per byte, low bits first, the high bit set on every byte but the last).

Bytes	Contents
4	AGID
1	format version, 1
1	number of tables, then per table:
1 + name	the name's length and the table's name in ASCII, as in the small database (imdb, tmdb_movie, ...)
4	the number of rows
one varint per row	the IDs in ascending order, each as its gap from the previous ID (the first from 0)
one byte per row	the ages, in the same order (0, 6, 12, 16 or 18)

pipeline/pack_ids.py writes it and has a reference reader (read).

The full database

Table	Contents
title	one row per Wikidata item: qid, label, year, kind (film, tv, other), fsk_official, fsk_estimate, probabilities p0 ... p18, confidence, had_article, source (official, batch, update), model, scored_at
external_id	pid, value, qid: every external ID Wikidata has for the title (IMDb is P345), 6.8 million in total
id_property	names of the external-ID properties
label	labels of linked items (people, genres, studios), used to build the model input for new titles
info	model name, decision threshold, build and update times

SELECT t.label, t.year, coalesce(t.fsk_official, t.fsk_estimate) AS age, t.confidence
FROM external_id e JOIN title t ON t.qid = e.qid
WHERE e.pid = 'P345' AND e.value = 'tt15239678';   -- Dune: Part Two, 2024, 12, NULL (official)

Weekly updates

Every Monday a scheduled job lists the films and TV shows on Wikidata, scores the ones that are not in the database yet with the same model, and publishes a new release if it added any. About 800 titles appear on Wikidata each week. The job does not refresh titles already in the database: an estimate stays even after Wikidata gets an official FSK rating for the title, and titles that are deleted or reclassified on Wikidata stay. A full rebuild picks those changes up. Details are in pipeline/README.md.

The current data

The databases were rebuilt from a Wikidata snapshot of 1 November 2026 with the v2 model: every item whose "instance of" is one of 16 film and TV classes (film, short film, animated film, television series, miniseries, anime, web series, ...).

	Films	TV	Other	Total
Titles	431,225	165,695	592	597,512
Official FSK rating	19,590	59	3	19,652
Estimate	411,635	165,636	589	577,860
With an IMDb ID	334,752	74,136	239	409,127
With a TMDB ID (movie or TV)	280,742	65,432	210	346,384
With a TheTVDB ID (series or movie)	200,938	52,453	172	253,563

Age	0	6	12	16	18
Official	1,741	3,134	7,935	5,516	1,326
Estimated	41,788	142,173	251,221	132,281	10,397

Titles without an English label fall back to the English or German Wikipedia title, then to a label in another language.

How the estimates are made

1. Facts: each title's Wikidata facts are written out as text, one line per fact: genres, country, director, cast, studio, network, series, awards, duration and so on. Ratings from other countries' boards are added (MPA PG-13, Kijkwijzer 12, ...), but never the FSK rating itself.
2. Wikipedia: when the title has an English or German Wikipedia article, its plot section or lead is appended. 36.8% of the estimated titles have one.
3. Model: LFM3-1.7B-Base with LoRA adapters and a classification head, fine-tuned on the 19,713 Wikidata titles that have an official FSK rating and on 1,381 TV shows whose FSK rating was looked up on German retail and streaming sites.
4. Decision: the estimate is the smallest age whose cumulative probability reaches 60%, so the model rounds up when it is unsure between two ages.

Titles with an official FSK rating keep it, and have no estimate.

How good the estimates are

On held-out titles (fold 0 of a 5-fold split), with the database's decision rule:

Held-out titles	Number	Exact	Within one step	Two or more steps too low
Films with an official FSK rating, with Wikipedia text	3,938	63.5%	97.3%	0.51%
the same films, Wikipedia text removed	3,938	61.4%	96.4%	0.71%
TV shows with an official FSK rating	59	57.6%	91.5%	1.69%
TV shows with a looked-up FSK rating	277	63.9%	98.6%	1.08%

• One step is one position on the scale 0 → 6 → 12 → 16 → 18.
• v2 against v1: the looked-up TV ratings raised TV accuracy from 51% to 64% exact on held-out shows and halved the suggestions that are too low (Game of Thrones, Shōgun and Succession now get 16 instead of 12). Films are unchanged on the test set. Retraining on the new snapshot moved about a quarter of the estimates for films without a Wikipedia article one step younger, whether or not the TV ratings were used; on held-out films with similarly sparse data, the models' suggestions were older than the official ratings on average.
• Label disagreement: rating boards often disagree by a step. FSK and the MPA give the same age to only 49.8% of the films they both rated.
• Obscure titles: titles with few Wikidata facts and no article are harder than the test set. New titles often start on Wikidata with only a few facts.

Details are in the model card.

What is here

Path	Contents
pipeline/	Python: Wikidata and Wikipedia collection, training, scoring, both database builds, the weekly update and release script
model/	model card, config.json and license files for the Hugging Face model repository edde746/agerating-lfm3-1.7b-v2, which also holds model.safetensors (66 MB)

pipeline/README.md lists every step of a full rebuild, from the Wikidata collection to the databases. The 1 November 2026 rebuild took about 3.5 hours: 1.5 hours of Wikidata and Wikipedia downloads on a Mac Studio, then 41 minutes of training, 79 minutes of scoring and a minute for the database on one RTX 4090.

License

The code, model weights and databases are under the agerating License 1.1:

• Free use: free and open-source projects may use them without asking, as long as the use is not commercial.
• Needs written approval: every commercial use, including commercial use by or in a free and open-source project, and every use in software that is not free and open source. Contact edde746 via GitHub.

The model weights are also subject to the LFM Open License v1.0 of the base model. The Wikidata facts are CC0 and can be taken from Wikidata directly. See NOTICE for all third-party terms.