# MSME Legal Dispute Classifier (Longformer, 6-Class)

## Model Overview
This model is a multi-class legal document classifier designed to categorize MSME-related dispute cases into six statutory dispute categories. It is fine-tuned from `allenai/longformer-base-4096` and optimized for long-form legal documents up to 1200 tokens. The system is intended for automated dispute categorization, legal triage, and decision-support applications in MSME dispute resolution workflows.

## Problem Statement
MSME dispute cases often involve lengthy legal narratives including:

- Statement of claim
- Buyer response
- Case summary
- Contractual and payment details

Manual classification is time-consuming and error-prone. This model automates dispute categorization into predefined legal classes.

## Classification Labels
The model predicts one of the following six categories:

| Label ID | Category                       |
|----------|--------------------------------|
| 0        | Delayed payment (no dispute)   |
| 1        | Quality dispute                |
| 2        | No formal contract             |
| 3        | Partial payment dispute        |
| 4        | Government procurement delay   |
| 5        | Service-related dispute        |

Label mapping is included in `label_mapping.json`.

## Model Architecture
- **Base Model**: Longformer
- **Checkpoint**: `allenai/longformer-base-4096`
- **Max Sequence Length**: 1200 tokens
- **Hidden Size**: 768
- **Number of Layers**: 12
- **Attention Type**: Local attention (CLS token classification)
- **Classification Head**: Linear layer (6 outputs)

Longformer was selected due to the long-document nature of legal dispute texts.

## Dataset Information
- Final Dataset Size (after cleaning): 2152 samples
- Duplicates Removed
- Label conflicts resolved
- Stratified 80–20 train/test split
- 5-fold stratified cross-validation

Class imbalance handled using weighted cross-entropy loss.

## Training Configuration
- **Optimizer**: AdamW
- **Learning Rate**: 2e-5
- **Batch Size**: 2
- **Gradient Accumulation Steps**: 4
- **Effective Batch Size**: 8
- **Epochs**: 3
- **Warmup Steps**: 200
- **Mixed Precision (FP16)**: Enabled
- **Loss Function**: Weighted Cross Entropy

## Evaluation Results (Held-Out Test Set)
Test Set Size: 431 samples

| Metric                  | Score |
|-------------------------|-------|
| Accuracy                | 0.77  |
| Macro Precision         | 0.76  |
| Macro Recall            | 0.74  |
| Macro F1 Score          | 0.75  |
| Macro AUC-ROC (OvR)     | 0.948 |

These results indicate strong class separability and balanced performance across all categories.

## Confusion Behavior
- Strong performance on Service-related and Government delay cases
- Moderate confusion between Quality disputes and Partial payment disputes
- Balanced macro performance across all classes

## Intended Use
This model is suitable for:
- Automated legal dispute classification
- MSME case triage systems
- Online Dispute Resolution (ODR) platforms
- Legal analytics systems
- Case routing and prioritization tools

## Limitations
- Performance may degrade for documents significantly exceeding 1200 tokens.
- Domain-specific to MSME dispute scenarios.
- Not designed for general legal classification tasks.
- Should not be used as a substitute for legal judgment.

## Ethical Considerations
This model is intended as a decision-support tool. Human oversight is recommended for legal decision-making applications. It does not provide legal advice.

## Usage Example
```python
from transformers import LongformerForSequenceClassification, AutoTokenizer
import torch

model = LongformerForSequenceClassification.from_pretrained("YOUR_USERNAME/msme-legal-dispute-classifier-longformer")
tokenizer = AutoTokenizer.from_pretrained("YOUR_USERNAME/msme-legal-dispute-classifier-longformer")

text = "The buyer failed to release payment within the agreed 45-day period."
inputs = tokenizer(text, truncation=True, max_length=1200, return_tensors="pt")
outputs = model(**inputs)
predicted_class = torch.argmax(outputs.logits, dim=1)
print("Predicted Label:", predicted_class.item())