promotion/Llama-3.1-8B-TLDR-Panacea-baseline
Llama-3.1-8B-TLDR-Panacea-baseline is an 8 billion parameter model developed by Panacea, based on the meta-llama/Llama-3.1-8B-Instruct backbone. It utilizes a DPO procedure with SVD-LoRA for preference optimization, specifically tuned for TL;DR objectives including coverage, faithfulness, conciseness, and helpfulness. This model is designed for generating high-quality summaries and concise responses, making it suitable for applications requiring efficient information extraction and presentation.
Loading preview...
Panacea on TL;DR
This model, Llama-3.1-8B-TLDR-Panacea-baseline, is an 8 billion parameter language model built upon the meta-llama/Llama-3.1-8B-Instruct backbone. It incorporates the Panacea method (Zhong et al., NeurIPS 2024) with a DPO (Direct Preference Optimization) procedure and linear scalarization. A key feature is its use of SVD-LoRA, where preference-agnostic singular values are combined with a preference vector, merged at an operating point determined on validation prompts.
Key Capabilities
- Optimized for TL;DR Objectives: Specifically tuned to excel in tasks requiring:
- Coverage: Ensuring all important points are included.
- Faithfulness: Maintaining accuracy relative to the source.
- Conciseness: Delivering information efficiently.
- Helpfulness: Providing useful and relevant summaries.
- Preference Optimization: Employs a sophisticated DPO and SVD-LoRA technique for fine-tuning based on preferences.
- Training Details: Underwent 300 optimizer updates with a global batch size of 16.
Evaluation
The model's performance is evaluated using an independent objective-wise win rate against a common reference, judged by Llama-3.3-70B-Instruct on prompt-disjoint held-out prompts. This evaluation protocol assesses its effectiveness across the specified TL;DR objectives.
Good For
- Generating concise summaries from longer texts.
- Applications requiring faithful and helpful information extraction.
- Use cases where brevity and accuracy are paramount.