license, base_model, tags, pipeline_tag, language

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

[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]