akashsukhija/complaint-bilstm-classifier

BiLSTM Customer Complaint Classifier

A Bidirectional LSTM model trained to classify consumer financial complaints into 36 sub-issue categories.

Model Details

Parameter	Value
Architecture	BiLSTM (2 layers, 256 hidden, bidirectional)
Embedding dim	128
Vocab size	30,522 (bert-base-uncased tokenizer)
Max sequence length	256
Total parameters	6,292,772
Dropout	0.3

Training

• Dataset: Johnade/consumer_complaints_cfpb (CFPB Consumer Finance Complaints)
• Classes: 36 sub-issue categories (filtered to classes with ≥500 examples)
• Optimizer: Adam (lr=1e-3, weight_decay=1e-5)
• Scheduler: ReduceLROnPlateau (patience=2, factor=0.5)
• Loss: CrossEntropyLoss with class weights for imbalanced data
• Epochs: 2
• Batch size: 16

Results

Metric	Value
Val Accuracy	0.1700
Val F1 (weighted)	0.1370
Best Epoch	1

Usage

import torch
from huggingface_hub import PyTorchModelHubMixin
import torch.nn as nn
from transformers import AutoTokenizer
import json

# Define the model class (needed for from_pretrained)
class BiLSTMClassifier(nn.Module, PyTorchModelHubMixin):
    def __init__(self, vocab_size=30522, embed_dim=128, hidden_dim=256,
                 num_layers=2, num_classes=36, dropout=0.3, pad_idx=0):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, embed_dim, padding_idx=pad_idx)
        self.bilstm = nn.LSTM(embed_dim, hidden_dim, num_layers, batch_first=True,
                              bidirectional=True, dropout=dropout if num_layers > 1 else 0.0)
        self.dropout = nn.Dropout(dropout)
        self.classifier = nn.Linear(hidden_dim * 2, num_classes)
    
    def forward(self, input_ids):
        x = self.dropout(self.embedding(input_ids))
        _, (hidden, _) = self.bilstm(x)
        hidden_cat = torch.cat([hidden[-2], hidden[-1]], dim=1)
        return self.classifier(self.dropout(hidden_cat))

# Load model and tokenizer
model = BiLSTMClassifier.from_pretrained("akashsukhija/complaint-bilstm-classifier")
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

# Load label mapping
from huggingface_hub import hf_hub_download
label_map = json.load(open(hf_hub_download("akashsukhija/complaint-bilstm-classifier", "label_mapping.json")))

# Classify a complaint
text = "My credit report shows an account that isn't mine. I've disputed it twice."
inputs = tokenizer(text, padding="max_length", truncation=True, max_length=256, return_tensors="pt")
with torch.no_grad():
    logits = model(inputs["input_ids"])
    pred = logits.argmax(dim=1).item()
print(f"Predicted: {label_map['id2label'][str(pred)]}")