valoricde / crewai-experiments

Experiments with local as well as models available through an api

CREWAI experiments

---

For my experiments with CrewAI, I decided to try 3 different projects, starting from the easiest to the most complex. The aim of the experiments was to have a team of AI agents do following work for me:

1. Examine my Startup idea
2. Build AI newsletter with Google SERP
3. Build AI newsletter with Reddit Scraper
4. Email classifier [WIP]

For my experiements I've tried following LLMs:

---

Available through API calls:

1. OpenAI -> GPT-4
2. Gemini Pro

Local Models through Ollama + rating of how they performed:

1. Mistral 7B

• Nice, coherent results
• Didn't understand that it should use scraping tool for the output
• Result is a bunch of generic text from training data

2. Mistral 7B instruct

• Nice, coherent results, a lot of emojis
• Didn't use any scraping tool for the output
• Result is a bunch of generic text from training data.

3. Open Chat 3.5 7B

• The best and most "newsletter-y" results
• But again, didn't use any tool, so generic content

4. Nous Hermes 7B

• Ok results
• didn't use any tool
• generic content

5. Open Hermes 2.5 7B

• The tone and style of writing is great
• but generic content
• didn't understand that it needs to use tools

6. Starling 7B

• Ok results
• didn't use any tool
• generic content

7. Llama 2 13B

• The only model that "understood" what the task is
• but the text wasn't coherent enough, didn's sound like a newsletter

8. Llama 2 13B chat

• Didn't understand the task or produce any output

9. Llama 2 13B text

• Didn't understand the task or produce any output

10. Llama 2 7B

• not coherent
• didn't use any tool
• no output

11. Llama 2 7B text

• No actual output
• didn't use any tool
• generic content

12. Llama 2 7B chat

• Didn't use any tool
• generic content

13. Phi-2

• The smallest model ran into biggest problems
• Lost track of what it's suppose to do, no output