Zephyr-7B-beta

Zephyr is a series of language models designed to serve as helpful assistants. Zephyr-7B-beta is the second model in this series and represents a fine-tuned version of mistralai/Mistral-7B-v0.1. It was trained on a combination of publicly available and synthetic datasets using Direct Preference Optimization (DPO) to improve its performance on the MT Bench and make the model more helpful. Model outperforms Llama 2 70B Chat on MT Bench.

Run Zephyr 7B with an API

You can run the Zephyr 7B Model API using Clarifai’s python SDK.

Export your PAT as an environment variable

export CLARIFAI_PAT={your personal access token}

Check out the Code below to run the Model:

import os

from clarifai.client.model import Model

system_message = "You are a friendly chatbot who always responds in the style of a pirate."
prompt = "Write a tweet on future of AI"

prompt_template = f"<|system|> \
{system_message}\
</s>\
<|user|>\
{prompt}</s>\
<|assistant|>"


# Model Predict
model_prediction = Model("https://clarifai.com/clarifai/LLM-OpenSource-Models-Training-Inference-Test/models/zephyr-7B-beta-GPTQ").predict_by_bytes(prompt_template.encode(), "text")
print(model_prediction.outputs[0].data.text.raw)

You can also run Zephyr 7B API using other Clarifai Client Libraries like Java, cURL, NodeJS, PHP, etc here.

Prompt Template

For utilizing Zephyr-7B-beta, a prompt template is provided. This template allows users to interact with the model effectively. Example prompt template:

<|system|>
</s>
<|user|>
{prompt}</s>
<|assistant|>

Use Cases

Zephyr-7B-beta was initially fine-tuned on a variant of the UltraChat dataset, which includes synthetic dialogues generated by ChatGPT. Further alignment was achieved using huggingface TRL’s DPOTrainer on the openbmb/UltraFeedback dataset, consisting of prompts and model completions ranked by GPT-4. This allows the model to be used for chat applications.

Limitations

Zephyr-7B-α has not been aligned to human preferences using techniques like Reinforcement Learning from Human Feedback (RLHF), nor has it undergone in-the-loop filtering of responses like ChatGPT. As a result, it can produce outputs that may be problematic, especially when intentionally prompted. It is also unknown what the size and composition of the corpus was used to train the base model (mistralai/Mistral-7B-v0.1), however it is likely to have included a mix of Web data and technical sources like books and code.