Text Generation
Transformers
Safetensors
English
mistral
conversational
text-generation-inference
polaris-small / README.md
ri-nat's picture
Update README.md
db84809 verified
|
raw
history blame
5.83 kB
metadata
license: apache-2.0
datasets:
  - StarfleetAI/function-calling
  - ajibawa-2023/Code-290k-ShareGPT
language:
  - en
library_name: transformers

StarfleetAI Polaris Small

Even though the model is not yet meant for production usage, we still encourage you to give it a try and tell us what you think.

Any feedback or suggestions are welcomed!

This model was developed by StarfleetAI with the aim of it being used for autonomous AI agent scenarios.

Model Details

  • Based on Open-Orca/Mistral-7B-OpenOrca
  • Additionally trained on coding tasks
  • Trained for function calling

Current Features

As of version 0.1 the model is:

  • Able to call functions
  • Not bad at coding tasks

Work To Do

  • Needs more training on examples where function call should be performed in the middle/end parts of the conversation, rather than in the beginning
  • Needs to eliminate some hallucinations (it, for example, sometimes regressing news articles instead of calling the get_rss_feed function, and stuff like that)
  • Correct typography (Answer: 123 instead of Answer:123)
  • Current date / time / timezone conversational abilities ("Schedule a call tomorrow at 6PM")
  • Needs some DPO
  • Need to be evaluated on relevant metrics

Prompt Format

We wanted our prompt to:

  1. Have a configurable system prompt part, which is meant to be changed by the end user.
  2. Describe the functions available to the model (in a specific, static place in the prompt, making it easier for the model to identify them).
  3. Be capable of carrying additional configuration options in the future, such as the current date/time or the user's name.

In order to fullfill these needs, we designed the prompt format, which seems to tick all the boxes.

With Functions

## Configuration

Functions: enabled

## Available Functions

{ ... }
{ ... }

---

You are a helpful assistant.

Without Functions

## Configuration

Functions: disabled

---

You are a helpful assistant.

Functions Calling

In order for the model to call a function, we have introduced two new tokens: <|fn_start|> and <|fn_end|>. It's safe to assume that if the model decides to call a function, there will be no other response from it except for the function call between these special tokens. If this is not true for some of your cases, please feel free to contact us with examples.

For now, the model is only capable of calling one function at a time.

Example Function Call

<|fn_start|>{"name": "generate_password", "arguments": {"length": 42}}<|fn_end|>

Model Description

  • Developed by: StarfleetAI
  • Model type: Mistral
  • Language(s) (NLP): English
  • License: Apache license 2.0
  • Finetuned from model: Open-Orca/Mistral-7B-OpenOrca

Training Details

You can find the trainer code available on Github under the Apache 2.0 license.