|
--- |
|
license: mit |
|
language: |
|
- am |
|
- ar |
|
- hy |
|
- eu |
|
- bn |
|
- bs |
|
- bg |
|
- my |
|
- hr |
|
- ca |
|
- cs |
|
- da |
|
- nl |
|
- en |
|
- et |
|
- fi |
|
- fr |
|
- ka |
|
- de |
|
- el |
|
- gu |
|
- ht |
|
- iw |
|
- hi |
|
- hu |
|
- is |
|
- in |
|
- it |
|
- ja |
|
- kn |
|
- km |
|
- ko |
|
- lo |
|
- lv |
|
- lt |
|
- ml |
|
- mr |
|
- ne |
|
- no |
|
- or |
|
- pa |
|
- ps |
|
- fa |
|
- pl |
|
- pt |
|
- ro |
|
- ru |
|
- sr |
|
- zh |
|
- sd |
|
- si |
|
- sk |
|
- sl |
|
- es |
|
- sv |
|
- tl |
|
- ta |
|
- te |
|
- th |
|
- tr |
|
- uk |
|
- ur |
|
- ug |
|
- vi |
|
- cy |
|
tags: |
|
- generated_from_trainer |
|
model-index: |
|
- name: verdict-classifier-en |
|
results: |
|
- task: |
|
type: text-classification |
|
name: Verdict Classification |
|
widget: |
|
- "One might think that this is true, but it's taken out of context." |
|
--- |
|
|
|
# Multilingual Verdict Classifier |
|
|
|
This model is a fine-tuned version of [xlm-roberta-base](https://huggingface.co/xlm-roberta-base) on 1,500 deduplicated multilingual verdicts from [Google Fact Check Tools API](https://developers.google.com/fact-check/tools/api/reference/rest/v1alpha1/claims/search), translated into 65 languages with the [Google Cloud Translation API](https://cloud.google.com/translate/docs/reference/rest/). |
|
It achieves the following results on the evaluation set, being 1,000 such verdicts, but here including duplicates to represent the true distribution: |
|
- Loss: 0.1856 |
|
- F1 Macro: 0.8148 |
|
- F1 Misinformation: 0.9764 |
|
- F1 Factual: 0.9375 |
|
- F1 Other: 0.5306 |
|
- Precision Macro: 0.8117 |
|
- Precision Misinformation: 0.9775 |
|
- Precision Factual: 0.9375 |
|
- Precision Other: 0.52 |
|
|
|
## Training procedure |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 2e-05 |
|
- train_batch_size: 4 |
|
- eval_batch_size: 4 |
|
- seed: 42 |
|
- gradient_accumulation_steps: 8 |
|
- total_train_batch_size: 32 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: linear |
|
- lr_scheduler_warmup_steps: 30066 |
|
- num_epochs: 1000 |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | F1 Macro | F1 Misinformation | F1 Factual | F1 Other | Precision Macro | Precision Misinformation | Precision Factual | Precision Other | |
|
|:-------------:|:-----:|:-----:|:---------------:|:--------:|:-----------------:|:----------:|:--------:|:----------:|:-------------------:|:------------:|:----------:| |
|
| 0.8707 | 1.0 | 3758 | 0.2414 | 0.7832 | 0.9639 | 0.7857 | 0.6 | 0.7950 | 0.9683 | 0.9167 | 0.5 | |
|
| 0.3918 | 2.0 | 7516 | 0.1856 | 0.8148 | 0.9764 | 0.9375 | 0.5306 | 0.8117 | 0.9775 | 0.9375 | 0.52 | |
|
| 0.1766 | 3.0 | 11274 | 0.1942 | 0.8394 | 0.9809 | 0.9538 | 0.5833 | 0.8349 | 0.9820 | 0.9394 | 0.5833 | |
|
| 0.1071 | 4.0 | 15032 | 0.2078 | 0.8676 | 0.9786 | 0.9841 | 0.64 | 0.8650 | 0.9797 | 1.0 | 0.6154 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.11.3 |
|
- Pytorch 1.9.0+cu102 |
|
- Datasets 1.9.0 |
|
- Tokenizers 0.10.2 |
|
|