---
license: cc-by-4.0
task_categories:
  - text-classification
language:
  - en
size_categories:
  - n<1K
pretty_name: IAB URL Gold 784
tags:
  - iab-taxonomy
  - url-classification
  - domain-classification
  - benchmark
  - multi-label
---

# IAB URL Gold 784

A **784-row gold benchmark for URL-only IAB classification**: given nothing but a bare
domain (rendered as a short text string), predict the
[IAB Tech Lab](https://iabtechlab.com/standards/content-taxonomy/) content and audience
categories of the web destination.

This is the held-out evaluation set used to benchmark
[`ZeroGPU/zlm-v1-iab-domain-classifier`](https://huggingface.co/ZeroGPU/zlm-v1-iab-domain-classifier)
against GPT-5.4-nano. On it, the fine-tuned 149M ModernBERT model reaches **content
micro-F1 0.3845** (P 0.4247 / R 0.3512) vs the nano baseline's **0.3526**
(P 0.3332 / R 0.3744), evaluated identically.

## Schema

One JSON object per line in `data/gold.jsonl`:

| Field | Type | Description |
| --- | --- | --- |
| `id` | string | Stable row id |
| `text` | string | The model input: `"<host> \| <segmented words> \| tld:<tld>"` (e.g. `"espn.com \| espn \| tld:com"`) |
| `content_labels` | list[string] | Gold IAB **Content** taxonomy labels |
| `audience_labels` | list[string] | Gold IAB **Audience** taxonomy labels (pipe-delimited paths) |
| `strat_tier1` | string | IAB content tier-1 category used for stratification |

## Provenance

- Domains were drawn from large public domain rankings, stratified across IAB content
  tier-1 categories.
- Gold labels were produced by **GPT-5.5** acting as an independent referee — a stronger
  model than either system evaluated on this set, and **not** the producer of the
  classifier's training labels, so the benchmark does not favor the fine-tuned model.
- All 784 rows were **excluded (by id and text) from the classifier's training corpus**.

## Evaluation protocol

Score **content micro-F1** against `content_labels`: for each row take the system's
predicted content labels (top-k = 6 for free-form LLM baselines; calibrated
score-threshold cut for the classifier), accumulate true/false positives and false
negatives over all rows, and compute precision, recall, and F1. Rows with empty
`content_labels` are skipped.

## Considerations

- Labels are model-generated (GPT-5.5), not human-annotated; treat them as high-quality
  silver-standard gold.
- The taxonomy includes sensitive categories (health, religion, adult content). The rows
  describe **web destinations**, not people.

## Citation

```bibtex
@misc{zerogpu2026iaburlgold784,
  title  = {IAB URL Gold 784: a gold benchmark for URL-only IAB classification},
  author = {ZeroGPU},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ZeroGPU/iab-url-gold-784}
}
```