SummLlama3.2-3B
| Property | Value |
|---|---|
| Parameter Count | 3.21B |
| Model Type | Summarization |
| Base Model | Llama-3.2-3B-Instruct |
| Paper | Research Paper |
| Tensor Type | BF16 |
What is SummLlama3.2-3B?
SummLlama3.2-3B is a specialized summarization model that has been fine-tuned from Llama-3.2-3B-Instruct using Direct Preference Optimization (DPO). The model was trained on over 100K summarization feedback instances spanning seven distinct domains, including news, lifestyle, medical, and various dialogue formats. This comprehensive training approach has resulted in significant improvements in summary quality across multiple dimensions.
Implementation Details
The model utilizes a sophisticated training methodology that leverages LLM-generated feedback instead of expensive human annotations. It achieves impressive metrics with faithfulness (0.867), completeness (0.598), and conciseness (0.686), showing substantial improvements over the base model.
- Optimized for both dialogue and non-dialogue content
- Implements Direct Preference Optimization (DPO)
- Trained across seven distinct domains
- Uses BF16 tensor format for efficient inference
Core Capabilities
- Generates highly faithful summaries that maintain factual accuracy
- Ensures comprehensive coverage of key information
- Produces concise outputs without redundant information
- Handles both conversational and formal text effectively
- Supports multi-domain summarization tasks
Frequently Asked Questions
Q: What makes this model unique?
This model stands out due to its optimization using DPO with large-scale summarization feedback, resulting in significantly improved performance across faithfulness, completeness, and conciseness metrics compared to the base Llama model.
Q: What are the recommended use cases?
The model is ideal for summarizing content across various domains, including news articles, medical texts, lifestyle content, and dialogue transcripts. It's particularly effective when high-quality, human-like summaries are required with strong emphasis on factual accuracy.
