README.md

---
pretty_name: Pandora 7B Chat
base_model: google/gemma-7b
datasets:
- danilopeixoto/pandora-instruct
- danilopeixoto/pandora-tool-calling
- danilopeixoto/pandora-rlhf
task_categories:
- text-generation
tags:
- chat
- dpo
- fine-tuning
- function-calling
- instruct
- rlhf
- sft
- tool-calling
license: bsd-3-clause
---

# Pandora 7B Chat

Pandora 7B Chat is a Large Language Model (LLM) designed for chat applications.

Pandora is fine-tuned with publicly available datasets, including a tool-calling dataset for agent-based tasks and a Reinforcement Learning from Human Feedback (RLHF) dataset with Direct Preference Optimization (DPO) training for preference alignment.

The fine-tuning process incorporates Low-Rank Adaptation (LoRA) with the [MLX framework](https://ml-explore.github.io/mlx/build/html/index.html), optimized for Apple Silicon.

The model is based on the [google/gemma-7b](https://huggingface.co/google/gemma-7b) model.

![Pandora](assets/pandora.jpeg)

## Datasets

Datasets used for fine-tuning stages:

- [danilopeixoto/pandora-instruct](https://huggingface.co/datasets/danilopeixoto/pandora-instruct)
- [danilopeixoto/pandora-tool-calling](https://huggingface.co/datasets/danilopeixoto/pandora-tool-calling)
- [danilopeixoto/pandora-rlhf](https://huggingface.co/datasets/danilopeixoto/pandora-rlhf)

## Evaluation

Evaluation on [MT-Bench](https://arxiv.org/abs/2306.05685) multi-turn benchmark:

![Benchmark](assets/benchmark.svg)

## Usage

Install package dependencies:

```shell
pip install mlx-lm
```

Generate response:

```python
from mlx_lm import load, generate


model, tokenizer = load('danilopeixoto/pandora-7b-chat')

prompt = '''<|start|>system
You are Pandora, a helpful AI assistant.
<|end|>
<|start|>user
Hello!
<|end|>
<|start|>'''

response = generate(model, tokenizer, prompt)
print(response)
```

The model supports the following prompt templates:

**Question-answering with system messages**

```txt
<|start|>system
{system_message}
<|end|>
<|start|>user
{user_message}
<|end|>
<|start|>assistant
{assistant_message}
<|end|>
```

**Tool calling**

```txt
<|start|>system
{system_message}
<|end|>
<|start|>system:tools
{system_tools_message}
<|end|>
<|start|>user
{user_message}
<|end|>
<|start|>assistant:tool_calls
{assistant_tool_calls_message}
<|end|>
<|start|>tool
{tool_message}
<|end|>
<|start|>assistant
{assistant_message}
<|end|>
```

> **Note** The variables `system_tools_message`, `assistant_tool_calls_message`, and `tool_message` must contain valid YAML.

An example of a tool-calling prompt:

```python
prompt = '''<|start|>system
You are Pandora, a helpful AI assistant.
<|end|>
<|start|>system:tools
- description: Get the current weather based on a given location.
 name: get_current_weather
 parameters:
 type: object
 properties:
 location:
 type: string
 description: The location name.
 required:
 - location
<|end|>
<|start|>user
What is the weather in Sydney, Australia?
<|end|>
<|start|>assistant:tool_calls
- name: get_current_weather
 arguments:
 location: Sydney, Australia
<|end|>
<|start|>tool
name: get_current_weather
content: 72°F
<|end|>
<|start|>'''
```

## Examples

**OpenGPTs**

![OpenGPTs](assets/opengpts.png)

## Copyright and license

Copyright (c) 2024, Danilo Peixoto Ferreira. All rights reserved.

Project developed under a [BSD-3-Clause license](LICENSE.md).

Gemma is provided under and subject to the [Gemma Terms of Use license](GEMMA_LICENSE.md).