Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,130 @@
+
+---
+---
+license: mit
+language:
+- it
+tags:
+- zero-shot-classification
+- text-classification
+- pytorch
+metrics:
+- accuracy
+- f1-score
+---
+# xlm-roberta-large-italian-legal-cap
+## Model description
+An `xlm-roberta-large` model finetuned on italian training data labelled 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
+```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)
+```
+
+#### Inference using the Trainer class
+```python
+model = AutoModelForSequenceClassification.from_pretrained('poltextlab/xlm-roberta-large-italian-legal-cap',
+                                                           num_labels=num_labels,
+                                                           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)
+
+```
+
+### Fine-tuning procedure
+`xlm-roberta-large-italian-legal-cap` was fine-tuned using the Hugging Face Trainer class with the following hyperparameters:
+```
+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=args.batch,
+    per_device_eval_batch_size=args.batch,
+    learning_rate=args.lr,
+    seed=42,
+    save_strategy='epoch',
+    evaluation_strategy='epoch',
+    save_total_limit=1,
+    load_best_model_at_end=True
+)
+```
+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 463 examples (10% of the available data).<br>
+Model accuracy is **0.82**.
+| label        |   precision |   recall |   f1-score |   support |
+|:-------------|------------:|---------:|-----------:|----------:|
+| 0            |        0.8  |     0.92 |       0.86 |        39 |
+| 1            |        0.5  |     0.29 |       0.36 |         7 |
+| 2            |        0.58 |     0.88 |       0.7  |         8 |
+| 3            |        0.87 |     0.87 |       0.87 |        23 |
+| 4            |        0.5  |     0.64 |       0.56 |        11 |
+| 5            |        0.88 |     0.88 |       0.88 |        26 |
+| 6            |        0.79 |     0.81 |       0.8  |        27 |
+| 7            |        0.85 |     0.92 |       0.88 |        12 |
+| 8            |        0.8  |     0.8  |       0.8  |         5 |
+| 9            |        0.86 |     0.9  |       0.88 |        41 |
+| 10           |        0.88 |     0.93 |       0.9  |        60 |
+| 11           |        0.83 |     0.45 |       0.59 |        11 |
+| 12           |        1    |     0.67 |       0.8  |         3 |
+| 13           |        0.86 |     0.8  |       0.83 |        40 |
+| 14           |        0.77 |     0.89 |       0.83 |        19 |
+| 15           |        0.94 |     0.94 |       0.94 |        16 |
+| 16           |        0.9  |     0.64 |       0.75 |        14 |
+| 17           |        0.88 |     0.72 |       0.79 |        39 |
+| 18           |        0.82 |     0.69 |       0.75 |        48 |
+| 19           |        0.38 |     0.75 |       0.5  |         4 |
+| 20           |        0.69 |     0.9  |       0.78 |        10 |
+| macro avg    |        0.78 |     0.78 |       0.76 |       463 |
+| weighted avg |        0.83 |     0.82 |       0.81 |       463 |
+
+## 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).
+
+## 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.
+
+If you encounter a `RuntimeError` when loading the model using the `from_pretrained()` method, adding `ignore_mismatched_sizes=True` should solve the issue.