Zephyr-7B-alpha is a brand new open-source language mannequin from HuggingFace and is predicated on Mistral-7B. This mannequin surpasses Llama 2 70B Chat on the MT Bench.
Now you can check out zephyr-7B-alpha within the Clarifai Platform and entry it by way of the API.
Desk of Contents
- Introduction
- Immediate Template
- Operating Zephyr 7B with Python
- Operating Zephyr 7B with Javascript
- Finest Use instances
- Limitations
Introduction
Zephyr-7B-alpha is the primary mannequin within the Zephyr sequence and is predicated on Mistral-7B. It has been fine-tuned utilizing Direct Choice Optimization (DPO) on a mixture of publicly accessible and artificial datasets. Notably, the in-built alignment of those datasets was eliminated to spice up efficiency on the MT Bench and make the mannequin extra useful.
Immediate Template
To work together successfully with the Zephyr-7B-alpha mannequin, use the immediate template beneath.
<|system|> system_prompt</s> <|consumer|> immediate</s> <|assistant|> |
This is an instance of the way to use the immediate template:
<|system|> |
Operating Zephyr 7B with Python
You’ll be able to run Zephyr 7B with our Python SDK with only a few traces of code.
To get began, Signup to Clarifai right here and get your Private Entry Token(PAT) underneath the safety part in settings.
Export your PAT as an surroundings variable:
export CLARIFAI_PAT=your private entry token |
Take a look at the Code Beneath:
Operating Zephyr 7B with Javascript
You too can run Zephyr Mannequin utilizing different Clarifai Consumer Libraries like Java, cURL, NodeJS, PHP, and so forth right here.
Mannequin Demo within the Clarifai Platform:
Check out the zephyr-7B-alpha mannequin right here: https://clarifai.com/huggingface-research/zephyr/fashions/zephyr-7B-alpha
Finest Use Instances
Chat purposes
The Zephyr-7B-alpha mannequin is well-suited for chat purposes. It was initially fine-tuned on a model of the UltraChat dataset, which incorporates artificial dialogues generated by ChatGPT. Additional refinement was achieved by using huggingface TRL’s DPOTrainer on the openbmb/UltraFeedback dataset. This dataset incorporates prompts and mannequin completions ranked by GPT-4. This intensive coaching course of ensures that the mannequin performs exceptionally nicely in chat purposes.
Limitations
Zephyr-7B-alpha has not been aligned to human preferences utilizing methods like Reinforcement Studying from Human Suggestions (RLHF). Because of this, it may possibly produce outputs that could be problematic, particularly when deliberately prompted.
Preserve in control with AI
-
Observe us on Twitter X to get the newest from the LLMs
-
Be a part of us in our Discord to speak LLMs