|
--- |
|
|
|
|
|
license: mit |
|
language: |
|
- cs |
|
--- |
|
# Model Card for small-e-czech-2stage-online-risks-cs |
|
|
|
<!-- Provide a quick summary of what the model is/does. --> |
|
|
|
This model is fine-tuned for 2nd stage multi-label text classification of Online Risks in Instant Messenger dialogs of Adolescents - it expects inputs where at least one of the classes appears. |
|
|
|
## Model Description |
|
|
|
The model was fine-tuned on a dataset of Instant Messenger dialogs of Adolescents. The classification is 2stage and the model outputs probablities for labels {0,1,2,3,4}: |
|
|
|
0. Aggression, Harassing, Hate |
|
1. Mental Health Problems |
|
2. Alcohol, Drugs |
|
3. Weight Loss, Diets |
|
4. Sexual Content |
|
|
|
- **Developed by:** Anonymous |
|
- **Language(s):** cs |
|
- **Finetuned from:** small-e-czech |
|
|
|
## Model Sources |
|
|
|
<!-- Provide the basic links for the model. --> |
|
|
|
- **Repository:** https://github.com/justtherightsize/supportive-interactions-and-risks |
|
- **Paper:** Stay tuned! |
|
|
|
## Usage |
|
Here is how to use this model to classify a context-window of a dialogue: |
|
|
|
```python |
|
import numpy as np |
|
import torch |
|
from transformers import AutoTokenizer, AutoModelForSequenceClassification |
|
|
|
# Prepare input texts. This model is pretrained on multi-lingual data |
|
# and fine-tuned on English |
|
test_texts = ['Utterance1;Utterance2;Utterance3'] |
|
|
|
# Load the model and tokenizer |
|
model = AutoModelForSequenceClassification.from_pretrained( |
|
'justtherightsize/small-e-czech-2stage-online-risks-cs', num_labels=5).to("cuda") |
|
|
|
tokenizer = AutoTokenizer.from_pretrained( |
|
'justtherightsize/small-e-czech-2stage-online-risks-cs', |
|
use_fast=False, truncation_side='left') |
|
assert tokenizer.truncation_side == 'left' |
|
|
|
# Define helper functions |
|
def predict_one(text: str, tok, mod, threshold=0.5): |
|
encoding = tok(text, return_tensors="pt", truncation=True, padding=True, |
|
max_length=256) |
|
encoding = {k: v.to(mod.device) for k, v in encoding.items()} |
|
outputs = mod(**encoding) |
|
logits = outputs.logits |
|
sigmoid = torch.nn.Sigmoid() |
|
probs = sigmoid(logits.squeeze().cpu()) |
|
predictions = np.zeros(probs.shape) |
|
predictions[np.where(probs >= threshold)] = 1 |
|
return predictions, probs |
|
|
|
def print_predictions(texts): |
|
preds = [predict_one(tt, tokenizer, model) for tt in texts] |
|
for c, p in preds: |
|
print(f'{c}: {p.tolist():.4f}') |
|
|
|
# Run the prediction |
|
print_predictions(test_texts) |
|
``` |
