Instructions to use LumeData/HandleAtlas-166m with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- GLiNER
How to use LumeData/HandleAtlas-166m with GLiNER:
from gliner import GLiNER model = GLiNER.from_pretrained("LumeData/HandleAtlas-166m") text = "Cristiano Ronaldo dos Santos Aveiro was born on 5 February 1985 in Funchal, Madeira, Portugal." labels = ["person", "date", "location"] entities = model.predict_entities(text, labels) for entity in entities: print(entity["text"], "=>", entity["label"]) - Notebooks
- Google Colab
- Kaggle
Add full model card
Browse files
README.md
CHANGED
|
@@ -1,3 +1,94 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
- multilingual
|
| 6 |
+
pipeline_tag: token-classification
|
| 7 |
+
tags:
|
| 8 |
+
- gliner
|
| 9 |
+
- ner
|
| 10 |
+
- token-classification
|
| 11 |
+
- social-media
|
| 12 |
+
- username-extraction
|
| 13 |
+
library_name: gliner
|
| 14 |
+
base_model: urchade/gliner_small-v2.1
|
| 15 |
---
|
| 16 |
+
|
| 17 |
+
# HandleAtlas-166m
|
| 18 |
+
|
| 19 |
+
A fine-tuned [GLiNER small v2.1](https://huggingface.co/urchade/gliner_small-v2.1) (~166M params)
|
| 20 |
+
for extracting social-media handles from short bios. Built on Twitter/X bios but the
|
| 21 |
+
patterns generalize to other platforms.
|
| 22 |
+
|
| 23 |
+
## Labels
|
| 24 |
+
|
| 25 |
+
- `instagram_username`
|
| 26 |
+
- `snapchat_username`
|
| 27 |
+
- `youtube_username`
|
| 28 |
+
- `twitch_username`
|
| 29 |
+
- `tiktok_username`
|
| 30 |
+
- `discord_username`
|
| 31 |
+
- `x_username`
|
| 32 |
+
- `cashapp_username`
|
| 33 |
+
- `onlyfans_username`
|
| 34 |
+
- `tumblr_username`
|
| 35 |
+
- `github_username`
|
| 36 |
+
- `kofi_username`
|
| 37 |
+
- `patreon_username`
|
| 38 |
+
- `roblox_username`
|
| 39 |
+
- `generic_username`
|
| 40 |
+
|
| 41 |
+
`generic_username` is a fallback for handle-shaped strings without a clear platform
|
| 42 |
+
indicator.
|
| 43 |
+
|
| 44 |
+
## Usage
|
| 45 |
+
|
| 46 |
+
```python
|
| 47 |
+
from gliner import GLiNER
|
| 48 |
+
|
| 49 |
+
model = GLiNER.from_pretrained("LumeData/HandleAtlas-166m")
|
| 50 |
+
|
| 51 |
+
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']
|
| 52 |
+
|
| 53 |
+
text = "Insta: foodgrammer | Snap: chefchef | DC: gamer420 | $cashtag"
|
| 54 |
+
for ent in model.predict_entities(text, labels, threshold=0.5):
|
| 55 |
+
print(f"{ent['text']!r} -> {ent['label']} ({ent['score']:.2f})")
|
| 56 |
+
```
|
| 57 |
+
|
| 58 |
+
## Training
|
| 59 |
+
|
| 60 |
+
- Base: `urchade/gliner_small-v2.1`
|
| 61 |
+
- Real data: ~1,000 hand-labeled Twitter bios
|
| 62 |
+
- Synthetic data: ~2,200 generated bios (template-based + IG→Discord text rewriting
|
| 63 |
+
for the discord_username class)
|
| 64 |
+
- Case augmentation: each training record is emitted in original + fully-lowercased
|
| 65 |
+
form so the model is robust to casing of platform prefixes (`Dc:`/`dc:`/`DC:` etc.)
|
| 66 |
+
- 5 epochs, batch 4 × grad-accum 2, lr 5e-6 (encoder) / 1e-5 (heads), cosine schedule
|
| 67 |
+
|
| 68 |
+
## Eval
|
| 69 |
+
|
| 70 |
+
On a 100-record held-out slice of real Twitter bios:
|
| 71 |
+
|
| 72 |
+
| metric | value |
|
| 73 |
+
|-----------|-------|
|
| 74 |
+
| precision | 0.849 |
|
| 75 |
+
| recall | 0.929 |
|
| 76 |
+
| F1 | 0.887 |
|
| 77 |
+
|
| 78 |
+
Strong per-label F1 on instagram (0.95), youtube (1.00), tiktok (1.00), twitch (1.00),
|
| 79 |
+
onlyfans (1.00), generic (0.88), cashapp (0.86), snapchat (0.80).
|
| 80 |
+
|
| 81 |
+
## Recommended thresholds
|
| 82 |
+
|
| 83 |
+
- Default: `threshold=0.5`
|
| 84 |
+
- For `generic_username`, bump to `0.65` to reduce false positives; it's the
|
| 85 |
+
catch-all label and over-fires at the default threshold.
|
| 86 |
+
|
| 87 |
+
## Limitations
|
| 88 |
+
|
| 89 |
+
- Trained on patterns common in Twitter/X bios; performance on other domains
|
| 90 |
+
(LinkedIn-style, Reddit, forum sigs) will be lower.
|
| 91 |
+
- `discord_invite` is not predicted — invite codes will be classified as
|
| 92 |
+
`discord_username` or skipped.
|
| 93 |
+
- Multi-line bios with many handles can occasionally confuse adjacent URL labels
|
| 94 |
+
(e.g., `patreon.com/x | github.com/x` chains).
|