JosefAlbers
commited on
Commit
•
15bf339
1
Parent(s):
3e06593
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,194 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
task_categories:
|
3 |
+
- question-answering
|
4 |
+
- translation
|
5 |
+
- summarization
|
6 |
+
- text-generation
|
7 |
+
- text2text-generation
|
8 |
+
- conversational
|
9 |
+
tags:
|
10 |
+
- agent
|
11 |
+
- multi-agent
|
12 |
+
- autogpt
|
13 |
+
- autogen
|
14 |
+
- agentgpt
|
15 |
+
- gptq
|
16 |
+
- wizard
|
17 |
+
- code-generation
|
18 |
+
- retrieval-augmented-generation
|
19 |
+
- humaneval
|
20 |
+
---
|
21 |
+
# [Roy: Rapid Prototyping of Agents with Hotswappable Components](https://github.com/JosefAlbers/Roy)
|
22 |
+
|
23 |
+
[<img src="https://colab.research.google.com/assets/colab-badge.svg" />](https://colab.research.google.com/github/JosefAlbers/Roy/blob/main/quickstart.ipynb)
|
24 |
+
[![DOI](https://zenodo.org/badge/699801819.svg)](https://zenodo.org/badge/latestdoi/699801819)
|
25 |
+
|
26 |
+
Roy is a lightweight alternative to `autogen` for developing advanced multi-agent systems using language models. It aims to simplify and democratize the development of emergent collective intelligence.
|
27 |
+
|
28 |
+
## Features
|
29 |
+
|
30 |
+
- **Model Agnostic**: Use any LLM, no external APIs required. Defaults to a 4-bit quantized wizard-coder-python model for efficiency.
|
31 |
+
|
32 |
+
- **Modular and Composable**: Roy decomposes agent interactions into reusable building blocks - templating, retrieving, generating, executing.
|
33 |
+
|
34 |
+
- **Transparent and Customizable**: Every method has a clear purpose. Easily swap out components or add new capabilities.
|
35 |
+
|
36 |
+
## Quickstart
|
37 |
+
|
38 |
+
```sh
|
39 |
+
git clone https://github.com/JosefAlbers/Roy
|
40 |
+
cd Roy
|
41 |
+
pip install -r requirements.txt
|
42 |
+
pip install -U transformers optimum accelerate auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
|
43 |
+
```
|
44 |
+
|
45 |
+
```python
|
46 |
+
from roy import Roy, Roys
|
47 |
+
roy = Roy()
|
48 |
+
s = '"What date is today? Which big tech stock has the largest year-to-date gain this year? How much is the gain?'
|
49 |
+
roy.generate(roy.format(s))
|
50 |
+
```
|
51 |
+
|
52 |
+
### **Rapid Benchmarking**
|
53 |
+
|
54 |
+
Roy provides a simple way to evaluate and iterate on your model architecture.. This allows you to:
|
55 |
+
|
56 |
+
- Easily swap out components, such as language models, prompt formats, agent architectures, etc
|
57 |
+
|
58 |
+
- Benchmark on different tasks like arithmetic, python coding, etc (default is OpenAI's HumanEval)
|
59 |
+
|
60 |
+
- Identify agent's areas of strengths and weaknesses
|
61 |
+
|
62 |
+
```python
|
63 |
+
from Roy.util import piecewise_human_eval
|
64 |
+
|
65 |
+
# Comparing different language models
|
66 |
+
piecewise_human_eval(0, lm_id='TheBloke/WizardCoder-Python-7B-V1.0-GPTQ')
|
67 |
+
# -> {'pass@1': 0.6341463414634146}
|
68 |
+
piecewise_human_eval(0, lm_id='TheBloke/tora-code-7B-v1.0-GPTQ')
|
69 |
+
# -> {'pass@1': 0.5609756097560976}
|
70 |
+
piecewise_human_eval(0, lm_id='TheBloke/Arithmo-Mistral-7B-GPTQ')
|
71 |
+
# -> {'pass@1': 0.5121951219512195}
|
72 |
+
|
73 |
+
# Testing a custom agent architecture
|
74 |
+
piecewise_human_eval(0, fx=<your_custom_Roy_agent>)
|
75 |
+
```
|
76 |
+
|
77 |
+
*Takes around 30 minutes each on a free Google Colab runtime.*
|
78 |
+
|
79 |
+
### **Constrained Beam Search**
|
80 |
+
|
81 |
+
Use templates to structure conversations (control output length, format, etc)
|
82 |
+
|
83 |
+
```python
|
84 |
+
roy.generate(s, ('\n```python', '\n```')) # Generate a python code block
|
85 |
+
roy.generate(s, (('\n```python', '\n```javascript'), '\n```')) # Generate python or javascript codes
|
86 |
+
roy.generate(s, ('\n```python', 100, '\n```')) # Generate a code block of size less than 100 tokens
|
87 |
+
```
|
88 |
+
|
89 |
+
### **Retrieval Augmented Generation**
|
90 |
+
|
91 |
+
Enhance generation with relevant knowledge.
|
92 |
+
|
93 |
+
```python
|
94 |
+
s = 'Create a text to image generator.'
|
95 |
+
r = roy.retrieve(s, n_topk=3, src='huggingface')
|
96 |
+
[roy.generate(s) for s in r]
|
97 |
+
```
|
98 |
+
|
99 |
+
### **Auto-Feedback**
|
100 |
+
|
101 |
+
Agents recursively improve via critiquing each other.
|
102 |
+
|
103 |
+
```python
|
104 |
+
s = "Create a secure and unique secret code word with a Python script that involves multiple steps to ensure the highest level of confidentiality and protection.\n"
|
105 |
+
for i in range(2):
|
106 |
+
c = roy.generate(s, prohibitions=['input'])
|
107 |
+
s += roy.execute(c)
|
108 |
+
```
|
109 |
+
|
110 |
+
### **Auto-Grinding**
|
111 |
+
|
112 |
+
Agents collaborate in tight loops to iteratively refine outputs to specification.
|
113 |
+
|
114 |
+
```python
|
115 |
+
user_request = "Compare the year-to-date gain for META and TESLA."
|
116 |
+
ai_response = roy.generate(user_request, ('\n```python', ' yfinance', '\n```'))
|
117 |
+
for i in range(2):
|
118 |
+
shell_execution = roy.execute(ai_response)
|
119 |
+
if 'ModuleNotFoundError' in shell_execution:
|
120 |
+
roy.execute(roy.generate(roy.format(f'Write a shell command to address the error encountered while running this Python code:\n\n{shell_execution}')))
|
121 |
+
elif 'Error' in shell_execution:
|
122 |
+
ai_response = roy.generate(roy.format(f'Modify the code to address the error encountered:\n\n{shell_execution}'))
|
123 |
+
else:
|
124 |
+
break
|
125 |
+
```
|
126 |
+
|
127 |
+
### **Multi-Agent**
|
128 |
+
|
129 |
+
Flexible primitives to build ecosystems of agents.
|
130 |
+
|
131 |
+
```python
|
132 |
+
roys = Roys()
|
133 |
+
|
134 |
+
# AutoFeedback
|
135 |
+
roys.create(agents = {'Coder': 'i = execute(generate(i))'})
|
136 |
+
roys.start(requests = {'i': 'Create a mobile application that can track the health of elderly people living alone in rural areas.'})
|
137 |
+
|
138 |
+
# Retrieval Augmented Generation
|
139 |
+
roys.create(
|
140 |
+
agents = {
|
141 |
+
'Retriever': 'r = retrieve(i)',
|
142 |
+
'Generator': 'o = generate(r)',
|
143 |
+
})
|
144 |
+
roys.start(requests = {'i': 'Create a Deutsch to English translator.'})
|
145 |
+
|
146 |
+
# Providing a custom tool to one of the agents using lambda
|
147 |
+
roys.create(
|
148 |
+
agents = {
|
149 |
+
'Coder': 'c = generate(i)',
|
150 |
+
'Proxy': 'c = custom(execute(c))',
|
151 |
+
},
|
152 |
+
tools = {'custom': lambda x:f'Modify the code to address the error encountered:\n\n{x}' if 'Error' in x else None})
|
153 |
+
roys.start(requests = {'i': 'Compare the year-to-date gain for META and TESLA.'})
|
154 |
+
|
155 |
+
# Another way to create a custom tool for agents
|
156 |
+
def custom_switch(self, c):
|
157 |
+
py_str = 'Modify the code to address the error encountered:\n\n'
|
158 |
+
sh_str = 'Write a shell command to address the error encountered while running this Python code:\n\n'
|
159 |
+
x = self.execute(c)
|
160 |
+
if 'ModuleNotFoundError' in x:
|
161 |
+
self.execute(self.generate(sh_str+x))
|
162 |
+
elif 'Error' in x:
|
163 |
+
self.dict_cache['i'] = [py_str+x]
|
164 |
+
else:
|
165 |
+
return '<<<Success>>>:\n\n'+x
|
166 |
+
|
167 |
+
roys.create(
|
168 |
+
agents = {
|
169 |
+
'Coder': 'c = generate(i)',
|
170 |
+
'Proxy': '_ = protocol(c)',
|
171 |
+
},
|
172 |
+
tools = {'protocol': custom_switch})
|
173 |
+
roys.start(requests = {'i': 'Compare the year-to-date gain for META and TESLA.'})
|
174 |
+
```
|
175 |
+
|
176 |
+
## Emergent Multi-Agent Dynamics
|
177 |
+
|
178 |
+
Roy aims to facilitate the emergence of complex, adaptive multi-agent systems. It draws inspiration from biological and AI concepts to enable decentralized coordination and continual learning.
|
179 |
+
|
180 |
+
- **Survival of the Fittest** - Periodically evaluate and selectively retain high-performing agents based on accuracy, speed etc. Agents adapt through peer interactions.
|
181 |
+
|
182 |
+
- **Mixture of Experts** - Designate agent expertise, dynamically assemble specialist teams, and route tasks to optimal experts. Continuously refine and augment experts.
|
183 |
+
|
184 |
+
These mechanisms facilitate the emergence of capable, adaptive, and efficient agent collectives.
|
185 |
+
|
186 |
+
## Get Involved
|
187 |
+
|
188 |
+
Roy is under active development. We welcome contributions - feel free to open issues and PRs!
|
189 |
+
|
190 |
+
## Support the Project
|
191 |
+
|
192 |
+
If you found this project helpful or interesting and want to support more of these experiments, feel free to buy me a coffee!
|
193 |
+
|
194 |
+
<a href="https://www.buymeacoffee.com/albersj66a" target="_blank"><img src="https://cdn.buymeacoffee.com/buttons/default-orange.png" alt="Buy Me A Coffee" height="25" width="100"></a>
|