---
language: en
license: apache-2.0
pipeline_tag: summarization
base_model: nsi319/legal-led-base-16384
tags:
- legal
- summarization
- led
- longformer
- long-document
- indian-court
- court-judgment
- judgment-summarization
- abstractive-summarization
- finetuned
- legal-nlp
datasets:
- ai4bharat/IndicCourtSummaries
- custom/legal-judgment-summaries
metrics:
- rouge
- bertscore
---

# 📘 Legal LED – Court Judgment Summarizer (Long-Document)

**Fine-tuned version of NSI’s Legal LED for abstractive summarization of long judicial decisions, case law, and court judgments.**

This model fine-tunes **`nsi319/legal-led-base-16384`**, a legally adapted LED with a 16k token context window, using a mix of *Indian court judgment data* and custom curated summaries.  
It is optimized to handle factual background, issues, arguments, legal reasoning, and final orders — producing structured and accurate case summaries.

---

# 🧠 Base Model

This model extends:

👉 **[nsi319/legal-led-base-16384](https://huggingface.co/nsi319/legal-led-base-16384)**

The original base model is pretrained on large-scale legal corpora including:

- Supreme Court & High Court judgments  
- Legislation  
- Legal treatises  
- Court orders  
- Regulatory material  

---

# 📚 Fine-Tuning Datasets

### **1️⃣ ai4bharat/IndicCourtSummaries** *(Recommended dataset)*  
Contains Indian Supreme Court & High Court judgments with human-written headnotes.  

Includes sections like:  
- Fact summary  
- Legal issues  
- Arguments  
- Court reasoning  
- Final conclusion  

### **2️⃣ Custom Legal Judgment Summaries**  
Additional curated summaries cleaned manually.

### Document Length  
Judgments ranged from **8,000 to 40,000 tokens**, making LED’s long-context attention crucial.

---

# ⚙️ Training Configuration

| Setting | Value |
|--------|--------|
| Base model | nsi319/legal-led-base-16384 |
| Epochs | 4 |
| Batch size | 1 |
| Gradient accumulation | 8 |
| Learning rate | 1e-5 |
| Optimizer | AdamW |
| Weight decay | 0.01 |
| FP16 | Yes |
| Warmup steps | 500 |
| Max input length | 14,000 tokens |
| Max output length | 512 tokens |
| Attention | Global attention on first token |
| Scheduler | Linear |

Training was performed on **NVIDIA TESLA P100 GPU (16GB)**.

---

# 🧪 Evaluation Metrics

### **Training Progress**
| Epoch | Training Loss | Validation Loss |
|-------|----------------|------------------|
| 1 | 2.25 | 2.32 |
| 2 | 2.03 | 2.11 |
| 3 | 1.85 | 2.03 |
| 4 | 1.79 | 2.01 | 

### **ROUGE (Judgment Test Set)**  
| Metric | F1 |
|--------|------|
| ROUGE-1 | 0.4610 |
| ROUGE-2 | 0.2353 |
| ROUGE-L | 0.2584 |

### **BERTScore**  
| Metric | Score |
|--------|--------|
| Precision | ~0.86 |
| Recall | ~0.85 |
| F1 | ~0.86 |

BERTScore shows strong semantic preservation — crucial for legal tasks.

---

# 🏗️ Long-Document Summarization Pipeline

Since judgments exceed **20k–30k tokens**, this model uses:

### 1️⃣ Paragraph-aware adaptive chunking  
### 2️⃣ Sliding window segmentation  
### 3️⃣ Chunk-wise LED summarization  
### 4️⃣ Top-K reranking using BERTScore  
### 5️⃣ Final LED rewrite pass  

This produces coherent, concise, and semantically accurate summaries.

---

# 📌 Intended Use

Designed for:

- Legal judgment summarization  
- Case-law research tools  
- Document automation  
- Student case summaries  
- Legal AI assistance  
- Preprocessing for RAG/legal LLM tasks  

---

# ⚠️ Limitations

- English only  
- Requires chunking for >16k tokens  
- May condense minority arguments  
- Not suitable for citation extraction or legal reasoning verification  
- Not a substitute for legal advice  

---

# 🔧 Usage Example

```python
from transformers import AutoTokenizer, LEDForConditionalGeneration
import torch

model_name = "Anurag33Gaikwad/legal-led-judgment-summarizer-16384"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = LEDForConditionalGeneration.from_pretrained(model_name)

text = """Your long court judgment here..."""

inputs = tokenizer(
    text,
    return_tensors="pt",
    truncation=True,
    max_length=4096
)

global_attention_mask = torch.zeros_like(inputs["input_ids"])
global_attention_mask[:, 0] = 1

summary_ids = model.generate(
    inputs["input_ids"],
    global_attention_mask=global_attention_mask,
    num_beams=5,
    max_length=512,
    early_stopping=True
)

print(tokenizer.decode(summary_ids[0], skip_special_tokens=True))