eval.yaml

## HandleAtlas Benchmark — evaluation spec
##
## Self-contained recipe for reproducing the numbers on the dataset card.
## Spans are character offsets into `text` (Python code points, end exclusive).
## Scoring is label-agnostic span F1 at IoU >= 0.5, plus exact span+label F1.

name: handleatlas-benchmark
version: 1
dataset:
  repo_id: LumeData/HandleAtlas-benchmark
  config: default
  split: test
  loader: |
    from datasets import load_dataset
    ds = load_dataset("LumeData/HandleAtlas-benchmark", split="test")

task:
  type: token-classification
  subtype: span-extraction
  input_field: text
  target_field: entities          # list[{start, end, label}]
  offsets: python-char            # not UTF-16, not BPE tokens
  end_exclusive: true
  multi_label_spans: true         # identical (start,end) with different labels are allowed

labels:
  - instagram_username
  - snapchat_username
  - youtube_username
  - twitch_username
  - tiktok_username
  - discord_username
  - x_username
  - cashapp_username
  - onlyfans_username
  - tumblr_username
  - github_username
  - kofi_username
  - patreon_username
  - roblox_username
  - generic_username

metrics:
  - id: span_only_f1
    description: Label-agnostic span detection. A prediction matches a gold span
      when their character-offset IoU >= iou_threshold. Each gold span is matched
      at most once (greedy by best IoU).
    iou_threshold: 0.5
    primary: true
    aggregate: micro
    reports: [precision, recall, f1, tp, fp, fn]

  - id: span_label_f1
    description: Exact match on (start, end, label). Stricter than span_only_f1
      and only fair to compare across models that share this label taxonomy.
    aggregate: micro
    reports: [precision, recall, f1, tp, fp, fn]

  - id: latency_ms
    description: Wall-clock ms per record, single batch=1 inference. CPU only.
    reports: [mean, p50, p95]
    hardware: MacBook Pro M5 Pro
    threads: 8

reference_implementation:
  language: python
  file: data/benchmark.py
  scoring_fns:
    - score_span_only
    - score_span_label

inference_protocol:
  threshold: 0.5
  per_label_threshold_overrides:
    generic_username: 0.65
  drop_predictions_with_label:
    - discord_invite           # excluded from this taxonomy
  decoding: greedy
  batch_size: 1

baselines:
  - model: LumeData/HandleAtlas-166m
    span_only_f1: 0.955
    span_label_f1: 0.887
    precision: 0.914
    recall: 1.000
    latency_ms_mean: 37.1
    latency_ms_p95: 52.2
    runtime: pytorch-float
  - model: LumeData/HandleAtlas-166m-CPU
    span_only_f1: 0.955
    span_label_f1: 0.887
    latency_ms_mean: 13.3
    latency_ms_p95: 22.3
    runtime: onnx-int8
  - model: urchade/gliner_small-v2.1
    span_only_f1: 0.061
    span_label_f1: 0.031
    precision: 1.000
    recall: 0.031
    latency_ms_mean: 35.0
    latency_ms_p95: 49.9
    runtime: pytorch-float
    note: zero-shot, same label list at threshold 0.5
  - model: openai/privacy-filter
    span_only_f1: 0.402
    precision: 0.305
    recall: 0.591
    latency_ms_mean: 113.6
    latency_ms_p95: 288.1
    runtime: pytorch-float
    note: different label taxonomy — span+label F1 not comparable

splits:
  test:
    n_records: 100
    n_spans: 127
    seed: 123
    notes: shuffle(annotations) -> take first 100; reproducible with SEED=123