---
license: apache-2.0
base_model: Qwen/Qwen2.5-3B
tags:
- legal
- legal-tech
- legal-clause-analysis
- lora
pipeline_tag: text-generation
language:
- en
---

# qwen2.5-3b-legal-merged

This repository contains a fine-tuned, merged version of the **Qwen2.5-3B** architecture, specifically optimized for legal domain text processing, contract deconstruction, and plain-English summary generation.

The model consists of the base architecture integrated with custom LoRA adapters trained on a curated dataset of legal clauses, terms of service agreements, and privacy policies.

## Model Description

- **Developed by:** Ayush Ghatak
- **Model Type:** Large Language Model (Causal LM)
- **Base Model:** Qwen/Qwen2.5-3B
- **Language(s):** English (Legal domain specialist)
- **License:** Apache-2.0
- **Fine-Tuning Method:** LoRA (Low-Rank Adaptation) merged directly into FP16 half-precision weights.

### Primary Use Cases
This model was designed and optimized to perform structured legal clause analysis split into twin streaming views:
1. **Plain-English Summarization:** Condensing complex boilerplate legalese into highly scannable, high-quality bullet points.
2. **Meta-Cognitive Reasoning:** Exposing an algorithmic, step-by-step cognitive roadmap identifying document type, legal parties, translating specific terms, and tracing logical breakdown.

---

## Intended Changes & Operational Triggers

For best performance in legal clause deconstruction tasks, prompt the model using structured phases. It has been trained to specifically respond to structural tags like `<Reasoning>:` and `<Summary>:` to allow clean UI-layer splitting:

### Expected Prompt/Response Layout
```text
[Input Text Block]

<Reasoning>:
1. IDENTIFY: [Document type and core legal purpose]
2. TRANSLATE: [Extraction and definition of domain-specific legal vocabulary]
3. COGNITIVE TRACE: [Step-by-step logic map of obligations and conditions]

<Summary>:
- [Core takeaway 1]
- [Core takeaway 2]