BERTweet-large-self-labeling

This model is a fine-tuned version of vinai/bertweet-large a dataset consisting of social media comments from 5 separate sources It achieves the following results on the evaluation set:

• Loss: 0.5607
• Accuracy: 0.7885
• F1 Macro: 0.7817
• F1 Weighted: 0.7885

Model description

We retrained the classification layer of Bert Base for a multi-label classification task on our self-labeled data. The model description of the base model can be found at the link above and the description of the dataset can be found here. The fine-tuning parameters are listed below. The initial model used in this experiment was bert-base-uncased. After decent results, we decided to use this model as it was pre-trained on a copious amount of Twitter data, which more closely aligned with our dataset. Turned out to be a good decision as this model was a 7.2% improvement over bert-base on the evaluation data.

Intended uses & limitations

Intended use for this model is to better understand the nature of different social media websites and the nature of the discourse on that site beyond the usual "positive", "negative", "neutral" sentiment of most models. The labels for the commentary data are as follows:

• Argumentative
• Opinion
• Informational
• Expressive
• Neutral

We think there is promise in this approach, and as this is the initial step towards a deeper understanding of social commentary, there are several limitations to outline

• As there were a total of 70k records, data was primarily labeled by language models, with the prompt including correctly labeled examples and incorrectly labeled examples with the correct label. Three language models were tasked with labeling, and only the majority vote labels were kept. Three-way tie samples were set aside. Future iterations would benefit from more models labeling, and more human labeled examples
• When reviewing records were ambiguous or that the classifier incorrectly predicted, it was clear that the labeling scheme is fuzzy in some instances. For instance, many "Opinion" comments can be viewed as "Expressive" "Arguments", leading to ambiguous labeling from models. It would be worth exploring a more nuanced labeling scheme, perhaps splitting "Expressive" into 2-3 labels and Opinion into another 1 or 2
• Due to the nature of the project, the commentary data used for training is subject to the following limitations
  • Queries were isolated to "politics" or "US politics"
  • All comment data is dated from Jan 1, 2025 to Feb 12, 2026, with the majority originating in 2026
  • We set a ceiling and a floor for number of comments per post. No posts with under 10 comments were used, and number of comments scraped were capped at 300

Training and evaluation data

A full description of the dataset can be found here

Training procedure

The full code used for training is below. We found overfitting to occur after 2 epochs

tokenizer = AutoTokenizer.from_pretrained("bert-base_uncased")

# Function to tokenize data with
def tokenize_function(batch):
    return tokenizer(
        batch['text'],
        truncation=True, 
        max_length=512 # Can't be greater than model max length
    )

# Tokenize Data
train_data = dataset['train'].map(tokenize_function, batched=True)
test_data = dataset['test'].map(tokenize_function, batched=True)
valid_data = dataset['valid'].map(tokenize_function, batched=True)

# Convert lists to tensors
train_data.set_format("torch", columns=['input_ids', "attention_mask", "label"])
test_data.set_format("torch", columns=['input_ids', "attention_mask", "label"])
valid_data.set_format("torch", columns=['input_ids', "attention_mask", "label"])

model = AutoModelForSequenceClassification.from_pretrained(
    MODEL_ID,
    num_labels=5, # adjust this based on number of labels you're training on
    device_map='cuda',
    dtype='auto',
    label2id=label2id,
    id2label=id2label
)

# Metric function for evaluation in Trainer
def compute_metrics(eval_pred):
    predictions, labels = eval_pred
    predictions = np.argmax(predictions, axis=1)
    
    return {
        'accuracy': accuracy_score(labels, predictions),
        'f1_macro': f1_score(labels, predictions, average='macro'),
        'f1_weighted': f1_score(labels, predictions, average='weighted')
    }

# Data collator to handle padding dynamically per batch
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

training_args = TrainingArguments(
    output_dir='./bert-comment',
    num_train_epochs=2,
    per_device_train_batch_size=32,
    per_device_eval_batch_size=64,
    learning_rate=2e-5,
    weight_decay=0.01,
    warmup_steps=300,
    
    # Evaluation & saving
    eval_strategy='epoch',
    save_strategy='epoch',
    load_best_model_at_end=True,
    metric_for_best_model='f1_macro',
    
    # Logging
    logging_steps=100,
    report_to='tensorboard',
    
    # Other
    seed=42,
    fp16=torch.cuda.is_available(),  # Mixed precision if GPU available
)

# Set up Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_data,
    eval_dataset=valid_data,
    processing_class=tokenizer,
    data_collator=data_collator,
    compute_metrics=compute_metrics
)

# Train!
trainer.train()

# Evaluate
eval_results = trainer.evaluate()
print(eval_results)

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 32
• eval_batch_size: 64
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 300
• num_epochs: 2
• mixed_precision_training: Native AMP

Training results

As this is a multi-label classification problem and there is class imbalance, the main metric we evaluate this model by is f1_macro

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1 Macro	F1 Weighted
0.5943	1.0	1540	0.5735	0.7708	0.7592	0.7708
0.3951	2.0	3080	0.5607	0.7885	0.7817	0.7885

Framework versions

• Transformers 5.0.0
• Pytorch 2.10.0+cu128
• Datasets 4.0.0
• Tokenizers 0.22.2