model card init

README.md

@@ -1,11 +1,9 @@
 
 ---
 ---
-license: mit
 language:
 - nl
 tags:
-- zero-shot-classification
 - text-classification
 - pytorch
 metrics:
@@ -14,83 +12,37 @@ metrics:
 ---
 # xlm-roberta-large-dutch-cap-v3
 ## Model description
-An `xlm-roberta-large` model finetuned on dutch training data labelled with [major topic codes](https://www.comparativeagendas.net/pages/master-codebook) from the [Comparative Agendas Project](https://www.comparativeagendas.net/).
+An `xlm-roberta-large` model fine-tuned on dutch training data labeled with [major topic codes](https://www.comparativeagendas.net/pages/master-codebook) from the [Comparative Agendas Project](https://www.comparativeagendas.net/).
 
 ## How to use the model
-#### Loading and tokenizing input data
+This snippet prints the three most probable labels and their corresponding softmax scores:
 ```python
-import pandas as pd
-import numpy as np
-from datasets import Dataset
-from transformers import (AutoModelForSequenceClassification, AutoTokenizer, 
-                          Trainer, TrainingArguments)
-
-CAP_NUM_DICT = {0: '1', 1: '2', 2: '3', 3: '4', 4: '5', 5: '6', 
-6: '7', 7: '8', 8: '9', 9: '10', 10: '12', 11: '13', 12: '14', 
-13: '15', 14: '16', 15: '17', 16: '18', 17: '19', 18: '20', 19: 
-'21', 20: '23', 21: '999'}
-
-tokenizer = AutoTokenizer.from_pretrained('xlm-roberta-large')
-num_labels = len(CAP_NUM_DICT)
-
-def tokenize_dataset(data : pd.DataFrame):
-    tokenized = tokenizer(data["text"],
-                          max_length=MAXLEN,
-                          truncation=True,
-                          padding="max_length")
-    return tokenized
-
-hg_data = Dataset.from_pandas(data)
-dataset = hg_data.map(tokenize_dataset, batched=True, remove_columns=hg_data.column_names)
-```
+import torch
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
 
-#### Inference using the Trainer class
-```python
-model = AutoModelForSequenceClassification.from_pretrained('poltextlab/xlm-roberta-large-dutch-cap-v3',
-                                                           num_labels=21,
-                                                           problem_type="multi_label_classification",
-                                                           ignore_mismatched_sizes=True
-                                                           )
-
-training_args = TrainingArguments(
-    output_dir='.',
-    per_device_train_batch_size=8,
-    per_device_eval_batch_size=8
-)
-
-trainer = Trainer(
-    model=model,
-    args=training_args
-)
-
-probs = trainer.predict(test_dataset=dataset).predictions
-predicted = pd.DataFrame(np.argmax(probs, axis=1)).replace({0: CAP_NUM_DICT}).rename(
-    columns={0: 'predicted'}).reset_index(drop=True)
+model = AutoModelForSequenceClassification.from_pretrained("poltextlab/xlm-roberta-large-dutch-cap-v3")
+tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-large")
 
-```
+sentence = "This is an example."
 
-### Fine-tuning procedure
-`xlm-roberta-large-dutch-cap-v3` was fine-tuned using the Hugging Face Trainer class with the following hyperparameters:
-```python
-training_args = TrainingArguments(
-    output_dir=f"../model/{model_dir}/tmp/",
-    logging_dir=f"../logs/{model_dir}/",
-    logging_strategy='epoch',
-    num_train_epochs=10,
-    per_device_train_batch_size=8,
-    per_device_eval_batch_size=8,
-    learning_rate=5e-06,
-    seed=42,
-    save_strategy='epoch',
-    evaluation_strategy='epoch',
-    save_total_limit=1,
-    load_best_model_at_end=True
-)
+inputs = tokenizer(sentence,
+                   return_tensors="pt",
+                   max_length=512,
+                   padding="do_not_pad",
+                   truncation=True
+                   )
+
+logits = model(**inputs).logits
+
+probs = torch.softmax(logits, dim=1).tolist()[0]
+probs = {model.config.id2label[index]: round(probability, 2) for index, probability in enumerate(probs)}
+top3_probs = dict(sorted(probs.items(), key=lambda item: item[1], reverse=True)[:3])
+
+print(top3_probs)
 ```
-We also incorporated an EarlyStoppingCallback in the process with a patience of 2 epochs.
 
 ## Model performance
-The model was evaluated on a test set of 6398 examples (10% of the available data).<br>
+The model was evaluated on a test set of 6398 examples.<br>
 Model accuracy is **0.83**.
 | label        |   precision |   recall |   f1-score |   support |
 |:-------------|------------:|---------:|-----------:|----------:|
@@ -118,13 +70,24 @@ Model accuracy is **0.83**.
 | macro avg    |        0.79 |     0.8  |       0.79 |      6398 |
 | weighted avg |        0.84 |     0.83 |       0.83 |      6398 |
 
+### Fine-tuning procedure
+This model was fine-tuned with the following key hyperparameters:
+
+- **Number of Training Epochs**: 10
+- **Batch Size**: 8
+- **Learning Rate**: 5e-06
+- **Early Stopping**: enabled with a patience of 2 epochs
+
 ## Inference platform
 This model is used by the [CAP Babel Machine](https://babel.poltextlab.com), an open-source and free natural language processing tool, designed to simplify and speed up projects for comparative research.  
 
 ## Cooperation
 Model performance can be significantly improved by extending our training sets. We appreciate every submission of CAP-coded corpora (of any domain and language) at poltextlab{at}poltextlab{dot}com or by using the [CAP Babel Machine](https://babel.poltextlab.com).
 
+## Reference
+Sebők, M., Máté, Á., Ring, O., Kovács, V., & Lehoczki, R. (2024). Leveraging Open Large Language Models for Multilingual Policy Topic Classification: The Babel Machine Approach. Social Science Computer Review, 0(0). https://doi.org/10.1177/08944393241259434
+
 ## Debugging and issues
-This architecture uses the `sentencepiece` tokenizer. In order to run the model before `transformers==4.27` you need to install it manually.
+This architecture uses the `sentencepiece` tokenizer. In order to use the model before `transformers==4.27` you need to install it manually.
 
 If you encounter a `RuntimeError` when loading the model using the `from_pretrained()` method, adding `ignore_mismatched_sizes=True` should solve the issue.