| --- |
| |
| license: apache-2.0 |
| base_model: "Qwen/Qwen2.5-0.5B-Instruct" |
| tags: ["legal-ai", "lora", "peft", "qlora", "indian-law", "legal-simplification", "accessibility", "irac", "transformers", "unsloth"] |
| language: ["en"] |
| pipeline_tag: text-generation |
| datasets: ["joyboseroy/inIRAC"] |
| library_name: transformers |
|
|
| --- |
| -------------------------- |
|
|
| # NyayaSaar-LoRA |
|
|
| NyayaSaar-LoRA is a lightweight LoRA adapter fine-tuned to simplify structured Indian legal reasoning into plain English. |
|
|
| The project focuses on improving accessibility to legal reasoning for non-lawyers interacting with the Indian legal system. |
|
|
| This model was trained using parameter-efficient fine-tuning (LoRA/QLoRA) on structured IRAC-style legal reasoning examples derived from the inIRAC dataset. |
|
|
| --- |
|
|
| # Motivation |
|
|
| Indian legal documents are often difficult for ordinary citizens to understand because of: |
|
|
| * archaic terminology |
| * procedural complexity |
| * dense sentence structures |
| * formal legal phrasing |
|
|
| NyayaSaar-LoRA explores whether small language models can learn to preserve legal meaning while simplifying language and improving readability. |
|
|
| The project prioritizes: |
|
|
| * accessibility |
| * explainability |
| * low-resource adaptation |
| * efficient fine-tuning |
|
|
| --- |
|
|
| # Example |
|
|
| ## Input |
|
|
| ```text id="jlwmkm" |
| Issue: Whether the detention order violates Article 22. |
| |
| Rule: Preventive detention laws require procedural safeguards. |
| |
| Application: The petitioner argued safeguards were not followed. |
| |
| Conclusion: The detention order is quashed. |
| ``` |
|
|
| ## Output |
|
|
| ```text id="4awft0" |
| The court examined whether the detention was legal under the Constitution. |
| |
| The law says preventive detention must follow proper safeguards. |
| |
| The petitioner argued these safeguards were ignored. |
| |
| The court agreed and cancelled the detention order. |
| ``` |
|
|
| --- |
|
|
| # Technical Details |
|
|
| ## Base Model |
|
|
| * Qwen2.5-0.5B-Instruct (4-bit quantized) |
|
|
| ## Fine-Tuning Method |
|
|
| * LoRA / QLoRA |
| * PEFT |
| * Unsloth |
| * Supervised Fine-Tuning (SFT) |
|
|
| ## Training Environment |
|
|
| * Google Colab |
| * Single GPU |
| * Low-resource fine-tuning setup |
|
|
| --- |
|
|
| # Dataset |
|
|
| Training examples were derived from the inIRAC dataset: |
|
|
| https://huggingface.co/datasets/joyboseroy/inIRAC |
|
|
| The dataset contains structured Indian legal reasoning in IRAC format: |
|
|
| * Issue |
| * Rule |
| * Application |
| * Conclusion |
|
|
| --- |
|
|
| # Intended Use |
|
|
| This project is intended for: |
|
|
| * legal accessibility research |
| * plain-English legal explanation |
| * educational demonstrations |
| * low-resource legal NLP experimentation |
| * research into explainable legal AI |
|
|
| --- |
|
|
| # Usage |
|
|
| ```python id="zj38zt" |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| from peft import PeftModel |
| |
| base_model = "Qwen/Qwen2.5-0.5B-Instruct" |
| adapter_model = "joyboseroy/nyayasaar-lora" |
| |
| tokenizer = AutoTokenizer.from_pretrained(base_model) |
| |
| model = AutoModelForCausalLM.from_pretrained( |
| base_model, |
| device_map="auto" |
| ) |
| |
| model = PeftModel.from_pretrained(model, adapter_model) |
| |
| prompt = """ |
| Issue: Whether the detention order violates Article 22. |
| |
| Rule: Preventive detention laws require procedural safeguards. |
| |
| Application: The petitioner argued safeguards were not followed. |
| |
| Conclusion: The detention order is quashed. |
| """ |
| |
| inputs = tokenizer(prompt, return_tensors="pt").to(model.device) |
| |
| outputs = model.generate( |
| **inputs, |
| max_new_tokens=200 |
| ) |
| |
| print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
| ``` |
|
|
| --- |
|
|
| # Limitations |
|
|
| This model: |
|
|
| * does not provide legal advice |
| * may oversimplify nuanced legal reasoning |
| * may omit procedural subtleties |
| * should not be used in high-stakes legal decision-making |
|
|
| Outputs must always be reviewed by qualified legal professionals before practical use. |
|
|
| --- |
|
|
| # Future Work |
|
|
| Potential future directions include: |
|
|
| * multilingual Indian legal simplification |
| * Hindi and Bengali adaptation |
| * graph-grounded legal reasoning |
| * retrieval-augmented legal explanation |
| * evaluation using readability and semantic-preservation metrics |
|
|
| --- |
|
|
| # Related Work |
|
|
| ## Dataset |
|
|
| inIRAC Dataset: |
| https://huggingface.co/datasets/joyboseroy/inIRAC |
|
|
| ## Research |
|
|
| Falkor-IRAC: Graph-Constrained Generation for Verified Legal Reasoning in Indian Judicial AI |
| https://arxiv.org/abs/2605.14665 |
|
|
| --- |
|
|
| # Citation |
|
|
| ```bibtex id="t1xjlwm" |
| @misc{bose2026nyayasaar, |
| title={NyayaSaar-LoRA: Simplifying Indian Legal Reasoning using PEFT}, |
| author={Joy Bose}, |
| year={2026}, |
| howpublished={Hugging Face Model Repository} |
| } |
| ``` |
|
|
| --- |
|
|
| # Author |
|
|
| Joy Bose |
| Senior Data Scientist and Researcher |
|
|
| Research interests: |
|
|
| * Legal AI |
| * Explainable AI |
| * Graph Reasoning |
| * Ethical AI |
| * Efficient LLM Adaptation |
|
|
| --- |
|
|
|
|
