Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
pipeline_tag: text-classification
|
| 4 |
+
tags:
|
| 5 |
+
- lfm2
|
| 6 |
+
- data-use
|
| 7 |
+
- gate
|
| 8 |
+
- domain
|
| 9 |
+
- multi-task
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# datause-classifier-multitask
|
| 13 |
+
|
| 14 |
+
Fine-tune of `LiquidAI/LFM2.5-Encoder-230M` with two heads: a page `has_data` gate and a
|
| 15 |
+
document `teratopic` domain classifier.
|
| 16 |
+
|
| 17 |
+
## Labels
|
| 18 |
+
30 teratopic sector themes (see `encoder_labels.json`).
|
| 19 |
+
|
| 20 |
+
## Training
|
| 21 |
+
- base model: `LiquidAI/LFM2.5-Encoder-230M`
|
| 22 |
+
- epochs: 3
|
| 23 |
+
- learning rate: 2e-05
|
| 24 |
+
- max_length: 2048
|
| 25 |
+
- precision: bf16
|
| 26 |
+
|
| 27 |
+
## Evaluation (holdout)
|
| 28 |
+
|
| 29 |
+
### Gate (page has_data)
|
| 30 |
+
- threshold: 0.95
|
| 31 |
+
- precision: 0.9409
|
| 32 |
+
- recall: 0.8785
|
| 33 |
+
- F1: 0.9086
|
| 34 |
+
- % pages flagged negative: 0.4604
|
| 35 |
+
|
| 36 |
+
### Domain (teratopic)
|
| 37 |
+
- micro F1: 0.5245
|
| 38 |
+
- macro F1: 0.3940
|
| 39 |
+
- mean average precision: 0.5031
|
| 40 |
+
|