tomg-group-umd/zephyr-llama3-8b-sft-refusal-n-contrast-multiple-tokens
The tomg-group-umd/zephyr-llama3-8b-sft-refusal-n-contrast-multiple-tokens model is an 8 billion parameter language model, fine-tuned from Meta-Llama-3-8B. It specializes in handling refusals and contrasting information, utilizing multiple refusal tokens for five distinct categories. This model is optimized for scenarios requiring nuanced control over model responses, particularly in refusal and contrastive generation tasks.
Loading preview...
Overview
This model, developed by tomg-group-umd, is an 8 billion parameter language model built upon the meta-llama/Meta-Llama-3-8B architecture. It has been fine-tuned using a combination of UltraChat SFT data with a respond token, CoCoNoT refusals with a refuse token, and CoCoNoT's contrast data. A key distinguishing feature is its implementation of multiple refusal tokens, specifically designed for each of five distinct categories.
Key Capabilities
- Refusal Handling: Incorporates specific refusal tokens to manage and categorize model refusals.
- Contrastive Generation: Fine-tuned with contrast data to generate nuanced and contrasting responses.
- Fine-tuned from Llama 3: Leverages the strong base capabilities of the Llama 3 8B model.
Usage Recommendations
For optimal generation, users are advised to refer to the specific code examples provided in the associated repository, particularly within the coconot_eval folder. This guidance is crucial for effectively utilizing the model's unique refusal and contrastive token mechanisms.