This model is a Longformer Encoder-Decoder (LED) baseline adapted for the summarization of Indian judicial texts (High Court and Supreme Court judgments). It was trained as part of a comparative multi-dataset benchmark evaluating domain-adapted transformer models on complex, lengthy legal discourse.

Architecture: Longformer Encoder-Decoder (LED)
Task: Abstractive Summarization
Language: English (Formal Indian Legal/Judicial Syntax)

⚠️ Important Limitation: Hardware Constraints

As detailed in our research, this specific LED baseline model suffered from under-training and did not reach full convergence. Due to the massive attention window required by the LED architecture and our rigid hardware constraints (Google Colab Free Tier, single Tesla T4 15GB VRAM GPU), we were unable to process sufficient batch sizes or complete the optimal number of epochs without encountering Out-Of-Memory (OOM) failures.

It is provided here for absolute research transparency, reproducibility, and as a baseline comparison against our more successful, resource-efficient architectures like T5-Legal and Legal-Pegasus (MCS).

Evaluation Results

Tested on a rigorous benchmark of 500 verified Indian legal judgments:

ROUGE-1: 0.334
ROUGE-2: 0.164
ROUGE-L: 0.315
BERTScore: 0.775
METEOR: 0.298

How to Use

You can easily instantiate this model for inference using the standard transformers library.

python

from transformers import AutoTokenizer, LEDForConditionalGeneration

# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("Rak-shit/legal-summeriser-led")
model = LEDForConditionalGeneration.from_pretrained("Rak-shit/legal-summeriser-led")

# Sample Indian Legal Text
legal_text = "---
license: cc-by-4.0
task_categories:
- text-classification
- token-classification
language:
- en
tags:
- legal
- india
- high-court
- judicial-judgment
pretty_name: High Court of Judicature Judgment Sample
---

# Dataset Card: Indian Judicial Judgment (High Court)

## Technical Overview
This is a synthetic reproduction of a standard Indian High Court civil judgment. It is formatted to test **NER** (Lawyers, Judges, Acts, Sections), **Citation Extraction**, and **Summarization** of legal reasoning.

## Legal Text: Civil Writ Petition

**IN THE HIGH COURT OF JUDICATURE AT BOMBAY**
**ORDINARY ORIGINAL CIVIL JURISDICTION**

**WRIT PETITION NO. 1402 OF 2026**

**M/s. Bharat Infrastructure Pvt. Ltd.**
A Company incorporated under the Companies Act, 2013, 
having its registered office at Nariman Point, Mumbai.       ... **Petitioner**

*Versus*

**The Municipal Corporation of Greater Mumbai**
Through the Municipal Commissioner, 
Mahapalika Marg, Mumbai.                                  ... **Respondent**

---

**CORAM:** HON’BLE MR. JUSTICE ARVIND S. DESAI &  
HON’BLE MS. JUSTICE PRIYANKA K. MEHTA

**DATE:** FEBRUARY 27, 2026

---

### JUDGMENT (Per: Arvind S. Desai, J.):

1. **Rule.** Respondents waive service. By consent of the parties, the Writ Petition is taken up for final hearing and disposal at the admission stage itself.

2. **The Facts:** The Petitioner, a construction firm, seeks a Writ of Mandamus directing the Respondent-Corporation to issue a Full Occupation Certificate (OC) for the project situated at Plot No. 88, Worli Division. Despite the issuance of a Part-OC on October 12, 2024, the Respondent has withheld the final certificate citing an alleged violation of the *Development Control and Promotion Regulations for Greater Mumbai, 2034* (DCPR 2034).

3. **Submissions:** Mr. Rajesh Sharma, learned Senior Counsel appearing for the Petitioner, submits that all requisite premiums and development charges have been paid in full. He relies upon the decision of the Apex Court in *State of Maharashtra v. Reliance Industries Ltd. (2022) 10 SCC 45*, wherein it was held that once a building is completed as per sanctioned plans, the authority cannot withhold an OC on grounds not explicitly stated in the initial Intimation of Disapproval (IOD).

4. **Respondent's Contention:** Mr. Sanjay Gupta, learned counsel for the Municipal Corporation, vehemently opposes the petition. He contends that an unauthorized mezzanine floor was constructed in the 'B-Wing' of the subject building, which attracts penalty under Section 53(1) of the *Maharashtra Regional and Town Planning Act, 1966* (MRTP Act).

5. **Analysis:** We have perused the records and the site inspection report dated January 15, 2026. It is evident that the alleged violation pertains to a minor internal modification that is regularizable under the prevailing compounding policy of the State Government. The Respondent's refusal to grant a Full OC for the entire project due to a localized issue in one wing is, in our view, disproportionate and contrary to the principles of administrative law.

6. **Order:** (a) The Respondent-Corporation is directed to process the Petitioner's application for a Full Occupation Certificate within a period of four (4) weeks from the date of this order.
    (b) The Petitioner is directed to pay the compounding charges as per the demand note dated February 5, 2026, within one week.
    (c) Rule is made absolute in the above terms. No order as to costs.

**[ARVIND S. DESAI, J.]** **[PRIYANKA K. MEHTA, J.]**

---

## Metadata for NLP Tasks
| Entity Type | Example Value |
| :--- | :--- |
| **Court** | High Court of Judicature at Bombay |
| **Statute** | MRTP Act, 1966 |
| **Section** | Section 53(1) |
| **Case Citation** | (2022) 10 SCC 45 |
| **Petitioner** | M/s. Bharat Infrastructure Pvt. Ltd. |
| **Judges** | A. S. Desai, P. K. Mehta |

# Tokenize and Generate Summary
inputs = tokenizer(legal_text, return_tensors="pt", max_length=4096, truncation=True)
summary_ids = model.generate(inputs["input_ids"], max_length=250, min_length=100, length_penalty=2.0, num_beams=4)

summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)
print("Generated Summary:", summary)

AUTHOR

Rakshit Gupta