license: mit
language:
- fr
tags:
- zero-shot-classification
- text-classification
- pytorch
metrics:
- accuracy
- f1-score
xlm-roberta-large-french-cap-v3
Model description
An xlm-roberta-large
model finetuned on french training data labelled with major topic codes from the Comparative Agendas Project.
How to use the model
Loading and tokenizing input data
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
model = AutoModelForSequenceClassification.from_pretrained('poltextlab/xlm-roberta-large-french-cap-v3',
num_labels=22,
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-french-cap-v3
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=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
)
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 2280 examples (10% of the available data).
Model accuracy is 0.71.
label | precision | recall | f1-score | support |
---|---|---|---|---|
0 | 0.71 | 0.72 | 0.71 | 200 |
1 | 0.59 | 0.44 | 0.5 | 62 |
2 | 0.82 | 0.74 | 0.78 | 80 |
3 | 0.66 | 0.75 | 0.7 | 64 |
4 | 0.72 | 0.57 | 0.63 | 186 |
5 | 0.75 | 0.76 | 0.76 | 125 |
6 | 0.7 | 0.6 | 0.65 | 85 |
7 | 0.88 | 0.82 | 0.85 | 45 |
8 | 0.7 | 0.74 | 0.72 | 57 |
9 | 0.74 | 0.86 | 0.79 | 58 |
10 | 0.82 | 0.77 | 0.8 | 154 |
11 | 0.55 | 0.65 | 0.59 | 105 |
12 | 0.76 | 0.64 | 0.7 | 87 |
13 | 0.58 | 0.59 | 0.59 | 106 |
14 | 0.8 | 0.8 | 0.8 | 87 |
15 | 0.7 | 0.72 | 0.71 | 46 |
16 | 0.57 | 0.71 | 0.63 | 59 |
17 | 0.64 | 0.79 | 0.71 | 204 |
18 | 0.78 | 0.78 | 0.78 | 359 |
19 | 0 | 0 | 0 | 7 |
20 | 0.76 | 0.7 | 0.73 | 104 |
21 | 0 | 0 | 0 | 0 |
macro avg | 0.65 | 0.64 | 0.64 | 2280 |
weighted avg | 0.72 | 0.71 | 0.71 | 2280 |
Inference platform
This model is used by the CAP Babel Machine, 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.
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.