File size: 2,886 Bytes
52f2f5a 6964783 52f2f5a 6964783 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 | ---
license: apache-2.0
language:
- en
- multilingual
pipeline_tag: token-classification
tags:
- gliner
- ner
- token-classification
- social-media
- username-extraction
library_name: gliner
base_model: urchade/gliner_small-v2.1
---
# HandleAtlas-166m
A fine-tuned [GLiNER small v2.1](https://huggingface.co/urchade/gliner_small-v2.1) (~166M params)
for extracting social-media handles from short bios. Built on Twitter/X bios but the
patterns generalize to other platforms.
## 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.
## Usage
```python
from gliner import GLiNER
model = GLiNER.from_pretrained("LumeData/HandleAtlas-166m")
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']
text = "Insta: foodgrammer | Snap: chefchef | DC: gamer420 | $cashtag"
for ent in model.predict_entities(text, labels, threshold=0.5):
print(f"{ent['text']!r} -> {ent['label']} ({ent['score']:.2f})")
```
## Training
- Base: `urchade/gliner_small-v2.1`
- Real data: ~1,000 hand-labeled Twitter bios
- Synthetic data: ~2,200 generated bios (template-based + IG→Discord text rewriting
for the discord_username class)
- Case augmentation: each training record is emitted in original + fully-lowercased
form so the model is robust to casing of platform prefixes (`Dc:`/`dc:`/`DC:` etc.)
- 5 epochs, batch 4 × grad-accum 2, lr 5e-6 (encoder) / 1e-5 (heads), cosine schedule
## Eval
On a 100-record held-out slice of real Twitter bios:
| metric | value |
|-----------|-------|
| precision | 0.849 |
| recall | 0.929 |
| F1 | 0.887 |
Strong per-label F1 on instagram (0.95), youtube (1.00), tiktok (1.00), twitch (1.00),
onlyfans (1.00), generic (0.88), cashapp (0.86), snapchat (0.80).
## Recommended thresholds
- Default: `threshold=0.5`
- For `generic_username`, bump to `0.65` to reduce false positives; it's the
catch-all label and over-fires at the default threshold.
## Limitations
- Trained on patterns common in Twitter/X bios; performance on other domains
(LinkedIn-style, Reddit, forum sigs) will be lower.
- `discord_invite` is not predicted — invite codes will be classified as
`discord_username` or skipped.
- Multi-line bios with many handles can occasionally confuse adjacent URL labels
(e.g., `patreon.com/x | github.com/x` chains).
|