agentlans/Qwen2.5-1.5B-Refiner

Qwen 2.5 1.5B Instruct - Text Refinement Model

A lightweight model fine-tuned to simplify and refine text for improved readability. Optimized for converting complex sentences into concise, clear versions while preserving core meaning.

Key Features:

• Reduces sentence complexity
• Converts passive to active voice
• Removes redundant phrases
• Maintains coherence in outputs

Training Details

Dataset:
Processed samples from finewebedu-refinement, containing 9996 simplified excerpts of FineWeb-Edu content. Texts were refined using:

1. Llama3.1-LexiHermes-SuperStorm
2. Fine-tuned Dolphin3.0-Llama3.2-3B

Method:

• Chunked inputs (~2,000 Llama 3 tokens)
• Applied LoRA (Rank 16, alpha 32) with rSLoRA
• NEFTune (noise=5), dropout=0.2
• 3 epochs via LLaMA Factory

Usage

Input Format:

Edit:
{{YOUR_TEXT_HERE}}

Output Format:

<text>Simplified version</text>

Limitations

• May oversimplify technical jargon
• Can omit references/formatting
• Rare garbled outputs
• Inherits biases from FineWeb-Edu
• Struggles with code/math notation

Licence

Apache 2.0