LumeData/HandleAtlas-benchmark

HandleAtlas Benchmark

Hand-labeled NER evaluation set for extracting social-media handles from Twitter / X bios. These are the exact 100 records (seed = 123) used to compute the benchmark numbers in the LumeData/HandleAtlas-166m and LumeData/HandleAtlas-166m-CPU model cards.

Schema

Each record:

{
  "id": 2,
  "text": "🍑 Ig | pea_arunya",
  "entities": [
    {"start": 7, "end": 17, "label": "instagram_username"}
  ]
}

• text — the raw bio (UTF-8, may contain emojis, multi-language text, line breaks).
• entities — list of NER spans. start and end are Python character indices (code points, not UTF-16 code units) into text. end is exclusive — text[start:end] gives the surface form of the entity.
• Multiple labels on the same span are encoded as separate entries with identical start/end (e.g. one handle written both as "ig + sc" gets two entries).

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

generic_username is a fallback for handle-shaped strings without a clear platform indicator (e.g. @somebody on twitter where the user is the writer's non-Twitter alt).

Size

split	records	spans
test	100	137

Reproducing the benchmark numbers

Reference scores on this split:

model	span-only F1	span+label F1
HandleAtlas-166m / -CPU	0.955	0.887
GLiNER-small v2.1 (zero-shot)	0.061	0.031
OpenAI Privacy Filter (1.5B)	0.402	n/a (different labels)

Usage

from datasets import load_dataset
ds = load_dataset("LumeData/HandleAtlas-benchmark", split="test")
print(ds[0])
# {"id": ..., "text": "🍑 Ig | pea_arunya",
#  "entities": [{"start": 7, "end": 17, "label": "instagram_username"}]}

Provenance

Bios were sampled from a public Twitter/X user-description CSV and labeled by hand using a single annotator. Each bio was either fully labeled (at least one platform handle present) or skipped. Skipped records are not included — every record in this dataset has at least one entity.

Limitations & ethics

• All bios are public Twitter/X profile descriptions but contain handles that point to real accounts. Treat the dataset accordingly when training downstream models.
• Single-annotator dataset; no inter-annotator agreement was measured.
• Skewed toward the most common platforms (instagram, snapchat). Rare-label performance should be validated against your own data before deploying.