simocorbo/toxicthesis-gemini-3.5-flash-lstm-classification-2

LSTM - GEMINI-3.5-FLASH - Classification (2 classes)

Toxicity prediction model trained on the GEMINI-3.5-FLASH dataset.

Property	Value
Model	LSTM
Task	Classification (2 classes)
Dataset	gemini-3.5-flash
Framework	PyTorch / PyTorch Lightning

Class: LSTMModel

from src.models.lstm import LSTMModel

model = LSTMModel(
    input_dim: int = 300,        # Dimension of input embeddings
    hidden_dim: int = 128,       # LSTM hidden dimension
    num_layers: int = 2,         # Number of LSTM layers
    dropout: float = 0.3,        # Dropout probability
    bidirectional: bool = True,  # Use bidirectional LSTM
    num_classes: int = 2,        # 1=regression, 2=binary, 3+=multi-class
    loss_fn: str = 'auto',
    lr: float = 0.001,
    gradient_clip_norm: float = 1.0
)

Methods

Method	Description
forward(inputs)	Input: (batch, seq_len, input_dim). Returns raw logits.
compute_loss(logits, targets)	Computes loss based on num_classes.
bin_targets(targets)	Converts continuous [0,1] targets to class indices.
from_config(config)	Create model from config dict.
load_from_checkpoint(path)	Load model from checkpoint file.

Usage with ToxicThesis (Recommended)

# 1. Clone ToxicThesis repository
# git clone https://github.com/simo-corbo/ToxicThesis
# cd ToxicThesis && pip install -r requirements.txt

from huggingface_hub import hf_hub_download
import torch
import numpy as np

# 2. Download checkpoint
checkpoint_path = hf_hub_download(
    repo_id="simocorbo/toxicthesis-gemini-3.5-flash-lstm-classification-2",
    filename="checkpoints/best.pt"
)

# 3. Import and load model from ToxicThesis
from src.models.lstm import LSTMModel

# Load using the built-in class method
model = LSTMModel.load_from_checkpoint(checkpoint_path, map_location='cpu')
model.eval()

# 4. Load FastText for embeddings
from src.utils.fasttext_utils import load_fasttext_model
ft = load_fasttext_model('cc.en.300.bin')

# 5. Get predictions
def predict(text: str, max_len: int = 128) -> dict:
    tokens = text.lower().split()[:max_len]
    embeddings = [ft.get_word_vector(w) for w in tokens] or [np.zeros(300)]
    
    # Pad sequence
    while len(embeddings) < max_len:
        embeddings.append(np.zeros(300))
    
    x = torch.tensor(np.array(embeddings[:max_len]), dtype=torch.float32).unsqueeze(0)
    
    with torch.no_grad():
        logits = model(x)
        
        if model.num_classes == 1:
            score = torch.sigmoid(logits).item()
            return {'score': score}
        elif model.num_classes == 2:
            prob = torch.sigmoid(logits).item()
            return {'probability': prob, 'class': int(prob >= 0.5)}
        else:
            probs = torch.softmax(logits, dim=-1).squeeze().tolist()
            return {'probabilities': probs, 'class': int(np.argmax(probs))}

result = predict("Your text here")
print(result)

Score Interpretation

Output	Range	Meaning
probability	[0, 1]	Probability of being toxic (class 1).
class	0 or 1	0 = non-toxic, 1 = toxic.

Decision boundary: Class 1 if probability >= 0.5.

Files

File	Description
checkpoints/best.pt	Model checkpoint (best validation loss)
hparams.yaml	Hyperparameters used for training
train.csv	Training metrics per epoch
val.csv	Validation metrics per epoch
vocab_stanza_hybrid.pkl	Vocabulary (for tree-based models)

Installation

# Clone ToxicThesis for full model implementations
git clone https://github.com/simo-corbo/ToxicThesis
cd ToxicThesis
pip install -r requirements.txt

# Or install dependencies directly
pip install torch transformers huggingface_hub fasttext-wheel stanza

Citation

@software{toxicthesis2025,
  title={ToxicThesis},
  author={Corbo, Simone},
  year={2025},
  url={https://github.com/simo-corbo/ToxicThesis}
}