⚖️ LawSummBart — Legal Document Summarization (ETA + LoRA Fine-Tuned BART)

LawSummBart is a domain-adapted abstractive summarization model fine-tuned for Indian legal documents, built using a novel Extract-Then-Assign (ETA) dataset generation pipeline and LoRA parameter-efficient fine-tuning on top of BART-Large CNN.

The model is optimized to handle long court judgments and produce concise, high-quality summaries suitable for downstream legal research applications.

---

🧠 Model Details

Model Description

• Shared by: qwertyqwerty070806
• Model Type: Abstractive Summarization (Seq2Seq)
• Language: English (Indian Legal Domain)
• Fine-tuned From: facebook/bart-large-cnn
• Framework: 🤗 Transformers
• License: MIT

This model uses an expanded training dataset generated via seven extractive summarization techniques combined with ROUGE & BERTScore-based sentence assignment, enabling improved domain adaptation for lengthy legal text.

---

📂 Model Sources

• GitHub: https://github.com/qwerty070806/LawSummBart
• Model Hub: https://huggingface.co/qwertyqwerty070806/LawSummBart
• Base Model Paper: https://arxiv.org/abs/1910.13461 (BART)

---

🧰 Uses

✔️ Direct Use

• Abstractive summarization of:
  • Court judgments
  • Legal case descriptions
  • Regulatory documents
  • Legal orders

✔️ Downstream Use

• Legal research assistants
• Automated judgment summarization
• Compliance document preprocessing
• NLP pipelines for legal analytics

🚫 Out-of-Scope Uses

• Providing legal advice
• Predicting legal outcomes
• Multi-lingual summarization (English only)
• Summaries requiring external legal expertise

---

⚠️ Bias, Risks & Limitations

Limitations

• Max effective input length ≈ 1024 tokens (BART constraint)
• Domain-specific model; may underperform on non-legal text
• Some legal nuances may be compressed or omitted

Potential Bias

• Court documents may contain systemic or institutional bias
• Extractive heuristics could overweight certain portions of text

Recommendation

Use model outputs as assistive summaries, not authoritative interpretations.
Human verification is essential for professional use.

---

🚀 Getting Started

import warnings
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

warnings.filterwarnings("ignore")

tokenizer = AutoTokenizer.from_pretrained("qwertyqwerty070806/LawSummBart", trust_remote_code=True)
model = AutoModelForSeq2SeqLM.from_pretrained("qwertyqwerty070806/LawSummBart", trust_remote_code=True).to("cuda")

text = """On March 5, 2021, the Securities and Exchange Commission charged AT&T..."""

inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512).to("cuda")

summary_ids = model.generate(
    **inputs,
    max_length=220,
    min_length=80,
    num_beams=5,
    length_penalty=1.0,
    no_repeat_ngram_size=3,
    early_stopping=True
)

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

Default Generation Parameters

Parameter	Value	Description
max_length	220	Maximum summary length
min_length	80	Minimum summary length
num_beams	5	Beam search width
length_penalty	1.0	Penalizes long sequences
no_repeat_ngram_size	3	Prevents repeated phrases
early_stopping	True	Stops when all beams reach EOS

---

🏗️ Training Details

Training Data

Dataset generated using Extract-Then-Assign (ETA):

• Applied seven extractive summarizers to each legal document
• Mapped extractive chunks to summary sentences using:
  • ROUGE scores
  • BERTScore semantic similarity
• Dataset expanded by 7× through extractive augmentation

Preprocessing

• Normalized paragraphs & whitespace
• Removed boilerplate legal headers (case metadata, citations)
• Tokenized up to BART’s limit (1024 tokens)

Training Procedure

LoRA Configuration

LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.1,
    target_modules=["q_proj", "v_proj"],
    task_type="SEQ_2_SEQ_LM"
)

Training Arguments

TrainingArguments(
    num_train_epochs=10,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=2,
    learning_rate=1e-4,
    eval_strategy="epoch",
    save_strategy="epoch",
    fp16=True,
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss"
)

Why LoRA?

• Trains only 5–10M parameters (vs 406M full fine-tuning)
• Saves 60%+ GPU memory
• Enables training on free GPUs (Kaggle, Colab)
• Maintains strong performance while reducing compute cost

---

📊 Evaluation

Metrics

• ROUGE-1, ROUGE-2, ROUGE-L
• BERTScore (Precision, Recall, F1)

Results

Metric	Score
ROUGE-1	0.5723
ROUGE-2	0.3547
ROUGE-L	0.3169
BERTScore Precision	0.8453
BERTScore Recall	0.8535
BERTScore F1	0.8493

Summary

The model achieves strong performance on abstractive summarization of long-form legal text, showing improved coherence and domain fidelity compared to base BART.

---

🌱 Environmental Impact (Approx.)

• Hardware: NVIDIA T4 (Kaggle/Colab)
• Training Time: ~7–8 hours
• CO₂ Emissions: Low (consumer-grade GPU)

---

🔬 Technical Specifications

Model Architecture

• Transformer encoder–decoder (BART)
• Seq2Seq language modeling objective
• LoRA applied to:
  • q_proj
  • v_proj

Compute Infrastructure

• Hardware: NVIDIA T4 (16 GB)
• Software:
  • PyTorch
  • Transformers ≥ 4.53
  • PEFT (LoRA)
  • Accelerate
  • Datasets

---

📚 Citation

@misc{lawsummbart2024,
  title = {LawSummBart: Legal Summarization with Extract-Then-Assign and LoRA Fine-Tuned BART},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/qwertyqwerty070806/LawSummBart}},
  note = {Accessed: [DATE]}
}

---

📬 Contact

For issues or questions, please visit: 👉 https://github.com/qwerty070806/LawSummBart