sail/Zephyr-7B-DICE-Iter2

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kTool Calling:SupportedPublished:Jul 3, 2024License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

sail/Zephyr-7B-DICE-Iter2 is a 7 billion parameter GPT-like language model developed by sail, fine-tuned from HuggingFaceH4/zephyr-7b-beta. This model utilizes the Bootstrapping Language Models with DPO Implicit Rewards (DICE) method at its second iteration, focusing on improving performance through synthetic datasets. It primarily supports English and demonstrates enhanced win rates on the AlpacaEval Leaderboard compared to its base model and prior iteration, making it suitable for general language generation tasks where improved instruction following is beneficial.

Loading preview...

Overview

sail/Zephyr-7B-DICE-Iter2 is a 7 billion parameter GPT-like language model developed by sail, fine-tuned from the HuggingFaceH4/zephyr-7b-beta base model. It was created using the "Bootstrapping Language Models with DPO Implicit Rewards" (DICE) method, specifically at its second iteration. This approach leverages synthetic datasets to enhance the model's capabilities, primarily in English.

Key Capabilities

  • Improved Instruction Following: Achieves a higher win rate on the AlpacaEval Leaderboard (20.16%) compared to its predecessor, Zephyr-7B-DICE-Iter1 (17.67%), and the base Zephyr-7b-beta (10.71%).
  • DICE Method Application: Incorporates the DICE (DPO Implicit Rewards) technique, as detailed in the research paper "Bootstrapping Language Models with DPO Implicit Rewards" (arXiv:2406.09760), for fine-tuning.
  • MIT Licensed: Available under the permissive MIT license, allowing for broad use and modification.

Good For

  • General Language Generation: Suitable for a wide range of text generation tasks where improved response quality and instruction adherence are desired.
  • Research and Development: Ideal for researchers and developers exploring advanced fine-tuning techniques like DICE and evaluating their impact on model performance.
  • Applications Requiring Enhanced Instruction Following: Can be integrated into applications that benefit from models with better alignment to user instructions and preferences.

Finetunes
1 models