The case for Laya in AdTech classification tasks

A Rust classifier that assigns IAB taxonomy categories to ad creatives and web articles with Laya. It depends on lisa-engine from agnosticeng/lisa and runs the model in-process — no external binary, no Python.

Contents

Cargo.toml                        the classifier crate (lisa-engine git dep)
src/main.rs                       the classifier
scripts/classify.py               thin wrapper (slug CLI over the same binary)
BENCHMARK.md                      CPU and Metal timings
data/
  taxonomy/
    ad-product-taxonomy-2.0.json  IAB Ad Product Taxonomy 2.0 as a tree
    content-taxonomy-3.1.json     IAB Content Taxonomy 3.1 as a tree
  inputs/
    ad-products.json              5 ad-product sample inputs from the case
    articles.json                 5 article sample inputs from the case

Build

cargo build --release

Requires an Apple-Silicon Mac with Metal (the --device cpu fallback works without a GPU), and Rust 1.97+.

Usage

# content articles (default taxonomy; uses each article's English `abstract`)
cargo run -- --all
cargo run -- --id article-2
cargo run -- --state "A new electric SUV review"

# ad products
cargo run -- --all \
  --taxonomy data/taxonomy/ad-product-taxonomy-2.0.json \
  --inputs data/inputs/ad-products.json \
  --state-field product --subject "ad creative"

Output (one line per record; timings go to stderr):

[model] loaded in 0.041s
  [article-2] q1(wide, 37 opts) 0.119s | q2(short, 8 opts) 0.077s
[article-2] total 0.196s
article-2: Sports (0.90)

option	default	meaning
--taxonomy	data/taxonomy/content-taxonomy-3.1.json	taxonomy JSON (nested tree)
--inputs	data/inputs/articles.json	records: {id, text, abstract?}
--model	convaiinnovations/laya	Laya checkpoint (repo id or dir)
--device	metal	metal or cpu
--shortlist	per taxonomy (8 content, 3 ads)	stage-2 option count
--stages	2	1 = wide choice only, 2 = wide + shortlist refine
--head-max-len	512	option-prompt token budget
--max-len	8192	total token budget
--depth	1	1 = top level only, 0 = descend to leaves
--read	auto	state source: auto (abstract then text), text, or abstract
--state-field / --subject	article / web article	state key and question wording
--temperature-by-options	—	TYPE:SIZE=TEMP, e.g. choice:11+=0.8
--id / --all / --state	—	select records, or pass ad-hoc text
--json	—	full JSON results (includes per-record timings)

scripts/classify.py is a thin wrapper with a slug CLI and per-taxonomy defaults; it calls the same binary (--classifier-bin to override, otherwise target/release, target/debug, then cargo run):

python scripts/classify.py --taxonomy content-taxonomy-3.1 --all
python scripts/classify.py --taxonomy ad-product-taxonomy-2.0 --state "Luxury SUVs"

Strategy

A single choice over a 34-46-option taxonomy is above Laya's ~20-option sweet spot. The classifier refines it in two stages, one forward pass each:

1. Wide stage — one choice over every candidate at the level; keep the top-N (--shortlist) by probability.
2. Refine stage — one small choice over that shortlist.

Repeat per level to build a path (--depth 0). Both stages are single batched calls, so this is 2 questions per level and runs in ~0.2s/record on Metal.

Data

Each taxonomy is a nested tree: tree is the list of root nodes and every node embeds its children (leaves have no children key). id is the unique key used in the Laya choice criteria.

{
  "id": "ad-product-taxonomy-2.0",
  "name": "IAB Ad Product Taxonomy",
  "version": "2.0",
  "node_count": 583,
  "root_count": 46,
  "tree": [
    { "id": "1002", "name": "Alcohol",
      "children": [ { "id": "1003", "name": "Bars" }, { "id": "1004", "name": "Beer" } ] }
  ]
}

Taxonomy	Version	Nodes	Roots	Tiers	Source
Ad Product	2.0	583	46	3	IAB
Content	3.1	704	37	4	IAB

Inputs: data/inputs/articles.json / data/inputs/ad-products.json, each {id, text, abstract?, url?, language?}. Articles are classified from their English abstract (a clean summary beats raw article text); ad creatives from text.

Results

On the bundled samples, two stages (--stages 2):

article	label	conf
Rogue OpenAI agent 'infiltrated' Australian government website in world first	Technology & Computing	0.80
Fury v Joshua announced for 11 December in Cardiff	Sports	0.90
Indian billionaire's payments firm plots biggest London flotation in years	Business and Finance	0.94
Hurricane Polo, rapidly intensifying from strong El Niño patterns, rages off coast of Mexico	Disasters	0.82
« L'idée d'un livre écrit par l'IA me dégoûte »… L'affaire Thélyson Orélien vue depuis les librairies	Books and Literature	0.97

set	labels
content (5 articles)	5/5
ad products (5 creatives)	5/5

For comparison, stage 1 alone already scores 5/5 on content (the wide choice is decisive on these articles), but only 3/5 on ad products — the refine stage is what fixes ad-1 and ad-3. The shortlist sizes are tuned per taxonomy on these small samples — content needs 8, ads need 3 — so treat them as starting points, not constants.

Notes and limits

Pulled from Laya's own model card, relevant here:

• High-cardinality choices degrade past ~20 options. This is why the two-stage shortlist exists; raise --head-max-len for wide stages if needed.
• Non-English state hurts the English checkpoint (the ad-4 sample is French; it only classifies once translated to English). Use English state, or --model convaiinnovations/laya-multilingual for non-Latin scripts.
• Confidence ships over-confident. Refit a temperature per question type on your own data before trusting probabilities; --temperature-by-options calibrates but cannot change the chosen label.
• Long inputs need a larger --max-len (up to 8192 on the multilingual encoder).

Performance

See BENCHMARK.md. Summary: ~0.2s/record on Metal, ~4-5s/record on CPU (2 questions/record, model load is negligible).

References

• https://iabtechlab.com/standards/ad-product-taxonomy/
• https://github.com/InteractiveAdvertisingBureau/Taxonomies/blob/main/Ad%20Product%20Taxonomies/Ad%20Product%20Taxonomy%202.0.tsv
• https://iabtechlab.com/standards/content-taxonomy/
• https://github.com/InteractiveAdvertisingBureau/Taxonomies/blob/develop/Content%20Taxonomies/Content%20Taxonomy%203.1.tsv
• https://huggingface.co/convaiinnovations/laya