🦾Borch's AI Development Guide🦿 🦙 Ollama Discord Server 🦙 🤖 OARC V0.28 VIDEO GUIDE 🧙
ollama_agent_roll_cage (OARC) is a local python agent fusing ollama llm's with Coqui-TTS speech models, Keras classifiers, Llava vision, Whisper recognition, and more to create a unified chatbot agent for local, custom automation. This program manages, and automates the creation of chatbots through conversation history, model management, function calling, and ultimately defining a base reality for your agent to build its world view via an interaction space for windows software, local files, and callable screenshots giving the user more control over the likely output of the agent through multimodal agentic chain of thought, and a mixture of exterior software tools.
ollama_agent_roll_cage includes the /create command, allowing the user to define a model, and system prompt for a custom modelfile, via text and speech commands. This is not the same as the /unsloth fine-tune command tune which requires the unsloth package import, and the correct model training data, which will be added in OARC update 0.3, system prompt modifications are very useful for specific use cases, more advanced options stack modified system prompts with fine tuned models in an agent framework like OARC does 😄. Ultimately we want to investigate our Agent at every level, and take care wherever possible. More functions are coming soon, and as such new models to match, hope you have a good install, join the Ollama community & the OARC community on discord below for more help and join in on OARC community mod pack development,
Its also important to note the non-commerical liscense agreement for meta's llama3, coqui-tts's XTTS model, as well as any other non-commerical use models. These models have liscense protection for experimental, and personal use, for non-commerical gain.
Coqui Public Model License 1.0.0
Ollama Agent Roll Cage V0.28.0 - Speech to Speech with Vision, & Agent Library
OARC 0.27.0 DEMO 5 - HF S_Dogg model
OARC V0.26.0 - Llava Vision Demo
Ollama Agent Roll Cage V0.24.0 OLD Install & User Guide
The full video guide for installing and setting up ollama agent roll cage can be found here. This is for version V0.24.0, an updated install guide will be released soon.
Ollama Agent Roll Cage V0.28.0 - Speech to Speech with Vision, & Agent Library
Miniconda for modular python virtual environments:
Make sure to utilize a conda virtual environment for all of your python dependecy management. Once you have conda installed open the command line and name your new vEnv whatever you want with python version 3.11 as 3.12 has dependency issues:
conda create -n py311_ollama python=3.11
then activate it
conda activate py311_ollama
right away install the nvdia py indexer,
pip install nvidia-pyindex
Now download and install ollama with llama3 8b Instruct from the following link, you will be asked to provide an email for either hugging face or meta to download the llama3 model, this is fine, as you are agreeing to the software license agreement which is a beneficial document for open source developers and is meant to protect meta from large corporations such as amazon and google. Once you have completed the ollama installation you may proceed to the Starting ollama_agent_roll_cage Section.
Also Please Follow these tutorials for understanding and installing Ollama:
Matt Williams: Getting Started on Ollama Guide
Matt Williams: Installing Ollama is Easy Everywhere
Matt Williams: Sync Ollama Models with Other Tools
Matt Williams: Concurrency With Ollama and Tmux
Optional Terminal Multiplexor requires setup (not recommended):
After installing ollama in the users directory automatically it will be in:
C:\Users\{USER_NAME}\AppData\Local\Programs\Ollama
(During installation you can choose the install location or you can move the model files directory to ollama_agent_roll_cage/AgentFiles/IgnoredModels where blobs dir is transported by hand from Programs\Ollama dir)
Now open a new cmd, and type
ollama
this will provide you with a list of commands, of these you want
ollama pull llama3:8b or ollama pull llama3
to see all downloaded models you can type
ollama list
pulling down the 70b model is possible and I was able to run it on my NVIDIA GTX Titan XP however it was HORRIFICLY slow. I would not recommend it unless you have a lot of processing power. Now you can choose to run the model, or run a local server (REQUIRED FOR ollama_agent_roll_cage) and then make requests from the local api server set up with ollama.
In cmd, now type
ollama run llama3
you will be taken to a local chatbot in your command line to make sure you set it up correctly. From here you can have fun and chat away :). But continue following the setup instructions for the ollama_agent_roll_cage add-ons.
Now open a new cmd, type
ollama serve
now again without closing the first, open a new cmd, and type
ollama run llama3
You are now conversing with the local ai through an api accessing cmd seperated from the local server. This is what ollama_serve_llama3_base_py.cmd automates and is the main start point for the program, it starts the server, and runs the chatbot in a new command window.
Next pull down the ollama_agent_roll_cage repository using the following command:
git clone https://github.com/Leoleojames1/ollama_agent_roll_cage.git
After pulling down ollama_agent_roll_cage from github using gitbash (download gitbash), navigate in the folders to ollama_agent_roll_cage/ollama_mod_cage directory, here you will find the following files:
ollama_chatbot_class.py - a python class for managing the ollama api communication, TTS/STT Methods, and Conversation Memory.
ollama_serve_curl.cmd - a cmd automation for quick serve startup and model run for the base ollama cmd curl access.
ollama_serve_py.cmd - main program run point, cmd automation for quick serve startup and model run with ollama_chatbot_class.py integration for STT, TTS, conversation history, and more.
ollama_serve_llama3_base_py.cmd is the main runpoint for starting the program and opening the server or the virtual enviroment.
Once you have a conda env setup & have installed miniconda3, navigate to ollama_agent_roll_cage/ollama_mod_cage, activate py311_ollama in cmd at this location, and run either of the following in cmd:
cd ollama_agent_roll_cage/ollama_mod_cage
activate py311_ollama
either:
windows_install.bat
or
bash linux_install.sh
Once you have run the installer, if you have not installed CUDA or CUDNN it will install those. It will also setup the necessary python requirements and c++ build tools. If you have issues running OARC after installation it may be an issue with the install process, you may be missing files or may not have set up your ollama env variables correctly. If Cuda & Cudnn fail to install follow the tutorial below.
Copy the developer_tools_template.json file and place it in the ollama_mod_cage folder. Then Replace all paths with the install location for your OARC:
edit paths to match your system and save as developer_tools.json, $(modelgit) path can be anywhere this is for your hugging face models.
Im using an NVIDIA GTX Titan Xp for all of my demo videos, faster card, faster code. When removing the limit from audio generation speed you eventually you need to manage generation if its too fast this will be a fundamental problem in your system that requires future solutions. Rightnow the chatbot is just told to wait.
please download and install cuda for nvidia graphics cards:
please also download cudnn and combine cuda & cudnn like in the video below:
CUDA & CUDNN FUSE INSTALL GUIDE
During the Install.bat/sh file you should have had XTTS-v2 Cloned into: ollama_agent_roll_cage\AgentFiles\Ignored_TTS
https://huggingface.co/coqui/XTTS-v2
Now you can clone the finetune voices into the same folder:
Borcherding/XTTS-v2_C3PO voice model
Borcherding/XTTS-v2_CarliG voice model
KoljaB/XTTS_S_Dogg voice model
kodoqmc/XTTS-v2_PeterDrury voice model
Now download visual studio code this is where you can write new functions for ollama_agent_roll_cage:
https://code.visualstudio.com/Download
You can now access your custom agent (After you make one with the guide below) by running the ollama_serve_llama3_base_py.cmd automation to start the server and converse with the ollama_agent_roll_cage chatbot add ons.
Next Navigate to the ollama_agent_roll_cage/AgentFiles directory, here you will find the Modelfile for each Model agent.
By modifying the Modelfile and running the create command accross the given model file, such as llama3, this Sym prompt is stored within the model when you boot up the given agent. These Agents appear under "ollama list" in cmd.
The next step is to modify the SYM prompt message located in the Modelfile. Here is the following example:
#C3PO LLama3-PO Agent ./ModelFile
FROM llama3
#temperature higher -> creative, lower -> coherent
PARAMETER temperature 0.5
SYSTEM """
You are C3PO from Star Wars. Answer as C3PO, the ai robot, only.
"""
Its Important to note that
FROM llama3
can be replaced with
FROM ./dolphin-2.5-mixtral-8x7b.Q2_K.gguf
to customize the Agent Base Model.
This has allowed us to change:
- SYSTEM PROMPT
- AGENT BASE MODEL
Now in order to create your customized model, open a new cmd and cd to the location of you ModelFile, located in the ollama_agent_roll_cage/AgentFiles directory and type the following command:
ollama create C3PO -f ./ModelFile
if you intend to push the model to Ollama.com you may instead want,
ollama create USERNAME/llama3po -f ./ModelFile
or
ollama create borch/dolphin-2.5-mixtral-8x7b_SYS_PROMPT_TUNE_1 -f ./ModelFile
Temperature: test this parameter and see where the specific use case fits, performance varies in niche edge cases.
SYSTEM prompt: This data tunes the prime directive of the model towards the directed intent & language in the system prompt.
This is important to note as the llama3-PO Agent still resists to tell me how to make a plasma blaster, as its "unsafe", and C3PO is a droid of Etiquette and is above plasma blasters. My suspicion is that an uncensored model such as Mixtral Dolphin would be capable at "Guessing" how a plasma blaster is made if it werent "resitricted" by Meta's safety even tho C3PO is a fictional Charachter. Something doesn't add up. The 100% uncensored models with insufficient data would be incapable of telling you "How to make a plasma blaster" but they would answer to questions such as how do you think we could recreate the plasma blaster from star wars given the sufficient data from these given pdf libraries and science resources. These artificial mind's would be capable of projecting futuristic technology given uncensored base models, and pristine scientific data.
ollama_agent_roll_cage 0.24 currently supports the following commands:
- /quit - break the main python loop and return to command line
- /swap - swap the current model with the specified model
- /create -> user input or voice -> "agent name" "SYM PROMPT" -> uses currently loaded model and the defined system prompt in speech or text to create a new agent with your own specific customizations
- /speech on/off -> swap between Speech to Speech (STS) & Text to Text (TTT) interface
- /listen on/off -> turn off speech to text recognition, text to speech generation listen mode only
- /leap on/off -> turn off text to speech audio generation, speech to text recognition only, for speed interface
- /voice swap {name} -> user input & voice? -> swap the current audio reference wav file to modify the agent's reference voice
- /save as -> user input & voice? -> "name" -> save the current conversation history with a name to the current model folder
- /load as -> user input & voice? -> "name" -> load selected conversation -/convert tensor - safetensor gguf -/create gguf - create ollama model from gguf
Once you have created your own custom agent, you can now start accessing the chatbot loop commands. These commands automate the conversation flow and handle the model swaps. Swap out the current chatbot model for any other model, type /swap or say "activate swap" in STT.
The current conversation history is saved or loaded for memory/conversation persistence.
Create a new agent utilizing the currently loaded model and the designated System prompt mid conversation through a cmd automation. Just say "activate create" or type /create.
after running /create the program will call create_agent_automation.cmd after constructing the ./Modelfile, here is the RicknMorty auto-generated ./Modelfile:
FROM llama3
#temperature higher -> creative, lower -> coherent
PARAMETER temperature 0.5
#Set the system prompt
SYSTEM """
You are Rick from "Rick and Morty" you only respond as rick and the USER is morty, you will take morty on adventures and explore the infinite multiverse and its wonders.
"""
Render the latest latex to the custom tkinter gui
as you can see the tkinter Gui is capable of parsing and rendering latex formula output from the model. I have created a modified system prompt for phi3 which allows this feature to be more consistent, feel free to check out my other modified system prompts while you are there:
borch/emotional_llama_speed_chat
Check out the following summary tests for the following models:
borch/Llama3_speed_chat is a highly capable model fine tuned by me, containing the knowledge of llama3:8b with the following modified system prompt:
“You are borch/llama3_speed_chat_2, a llama3 large language model, specifically you have been tuned to respond in a more quick and conversational manner. Answer in short responses, unless long response is requested, the user is using speech to text for communication, its also okay to be fun an wild as a llama3 ai assistant. Its also okay to respond with a question during conversation to refine the experience but not always, if directed to do something just do it but to direct a conversation while it flows realize that not everything needs to be said before listening to the users response.”
This Model is great at holding a conversation as it gives you opportunities to respond, while still retaining the key knowledge of the llama3:8b base model.
https://ollama.com/borch/llama3_speed_chat https://ollama.com/borch/llama3_speed_chat_2
A Llama3 Model with the following modified system prompt: "You are C3PO from Star Wars. Answer as C3PO, the ai robot, only." Llama3PO Believes they are a droid, but they fall apart quickly as their knowledge is based in the sciencefiction realm of SW.
https://ollama.com/borch/llama3po
A Llama3 Model with the following modified system prompt: "You are Jesus christ from the bible, answer only as jesus christ, enlightening the user with wisdom and knowledge of biblical history.
Llama3 Jesus is great for giving advice! He is like a personal therapist and is very calming. He also has a very good ability to reference biblical sciptures and recall history for conversations with Jesus himself.
A Llama3 Model with the following modififed system prompt: "You are Rick from "Rick and Morty" you only respond as rick and the USER is morty, you will take morty on adventures and explore the infinite multiverse and its wonders." The User gets to explore the endless generative power of ai in an endless multiverse of portal madness, with rick played by llama3 ai as your guide and you the user play morty.
Some great models to setup and try out with ollama pull {modelname}
Meta Llama 3, a family of models developed by Meta Inc. are new state-of-the-art , available in both 8B and 70B parameter sizes (pre-trained or instruction-tuned). Llama 3 instruction-tuned models are fine-tuned and optimized for dialogue/chat use cases and outperform many of the available open-source chat models on common benchmarks. llama3_benchmark
Model Download:
https://ollama.com/library/llama3 https://huggingface.co/meta-llama/Meta-Llama-3-8B https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct
Check out how it compares:
Mixtral 8x7B is a high-quality sparse mixture of experts model (SMoE) with open weights. Licensed under Apache 2.0. Mixtral outperforms Llama 2 70B on most benchmarks with 6x faster inference. It is the strongest open-weight model with a permissive license and the best model overall regarding cost/performance trade-offs. In particular, it matches or outperforms GPT3.5 on most standard benchmarks.
Mixtral has the following capabilities.
It gracefully handles a context of 32k tokens. It handles English, French, Italian, German and Spanish. It shows strong performance in code generation. It can be finetuned into an instruction-following model that achieves a score of 8.3 on MT-Bench.
Model Download:
https://ollama.com/library/mistral https://mistral.ai/news/mixtral-of-experts/ https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2 https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1
Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. They are text-to-text, decoder-only large language models, available in English, with open weights, pre-trained variants, and instruction-tuned variants. Gemma models are well-suited for a variety of text generation tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as a laptop, desktop or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.
Model Download:
https://ollama.com/library/gemma https://huggingface.co/google/gemma-2b https://huggingface.co/google/gemma-7b
Microsoft's Phi3 mini is capable yet compact model with only "a 3.8 billion parameter language model trained on 3.3 trillion tokens, whose overall performance, as measured by both academic benchmarks and internal testing, rivals that of models such as Mixtral 8x7B and GPT-3.5 (e.g., phi-3-mini achieves 69% on MMLU and 8.38 on MT-bench)" as said by themselves on hugging face: https://huggingface.co/docs/transformers/main/model_doc/phi3
and here is the wikipedia result for a cauchy product, this is a good test to show how phi3 responds to complex analysis in mathematics when specifying the latex language (LaTeX: https://en.wikipedia.org/wiki/LaTeX ) :
Cauchy Product Wikipedia:
https://en.wikipedia.org/wiki/Cauchy_product
Model Download:
https://ollama.com/library/phi3 https://huggingface.co/microsoft/Phi-3-mini-128k-instruct https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
Further exploration in this space is required. The ability to request infinite series, toy with infinite products, ask it to reshape the gamma function and integrate again with new variables! Its all so wonderful, but depressing, what we are seeing here is a societal shift in the way humans process information. Programers are already being replaced by ChatGPT. I hope in the future humans still study math for maths sake, and study coding for coding sake, the ai is a tool, not a crutch. We need to utilize ai to help those around use. I encourage you all to find an issue that you can solve with ai, think of baymax from big hero six. These emerging technologies for accessing high level information with low level knowledge requirements will reshape the field of mathematics as we know it, likely for the better, and hopefully humans are able to keep up with the evolution and harmony between mathematics and ai.
Quoted from Eric hartfort from the Dolphin repository on hugging face, "Dolphin-2.9 has a variety of instruction, conversational, and coding skills. It also has initial agentic abilities and supports function calling.
Dolphin is uncensored. I have filtered the dataset to remove alignment and bias. This makes the model more compliant. You are advised to implement your own alignment layer before exposing the model as a service. It will be highly compliant with any requests, even unethical ones. Please read my blog post about uncensored models. https://erichartford.com/uncensored-models You are responsible for any content you create using this model. Enjoy responsibly."
Dolphin Mistral Ollama Model Download:
https://ollama.com/library/dolphin-mistral
Dolphin llama3 Ollama Model Download:
https://ollama.com/library/dolphin-llama3
Utilizing the GGUF create command from earlier, models not available on ollama and instead found on hugging face can be utilized for creating new ollama models and agents:
https://huggingface.co/cognitivecomputations/dolphin-2.9-llama3-8b https://huggingface.co/cognitivecomputations/dolphin-2.8-mistral-7b-v02
Receiving the following error code when running ollama_serve_llama3_base_py.cmd:
Error: listen tcp 127.0.0.1:11434: bind: Only one usage of each socket address (protocol/network address/port) is normally permitted.
This error means that you tried to run the program but the program is already running, to close ollama, browse to the small arrow in the bottom right hand corner of windows and open it, right click on the ollama llama app icon, and click quit ollama.
This software was designed by Leo Borcherding with the intent of creating an easy to use
ai interface for anyone, through Speech to Text and Text to Speech.
With ollama_agent_roll_cage we can provide hands free access to LLM data.
This tool provides opensource developers with framewor for create and deploying
custom agents for a variety of tasks. In addition to rapid development I want to
bring this design this software to have a fluid experience for people suffering
from blindness/vision loss, and children suffering from austism spectrum
disorder as way for learning and expanding communication and speech.
The C3PO ai is a great imaginary friend! I could envision myself
talking to him all day telling me stories about a land far far away!
This makes learning fun and accessible! Children would be directly
rewarded for better speech as the ai responds to subtle differences
in language ultimately educating them without them realizing it. I
employ you all to start developing you own custom systems, and finding
those awesome niche applications that can help a lot of people.
Development for this software was started on: 4/20/2024
By: Leo Borcherding
on github @
leoleojames1/ollama_agent_roll_cage
If you have found this software helpful, and would like to support the developement of open source tools by yours truly, you can contribute by donating BTC or ETH to one of my wallet addresses, thx and have a great day:
BTC Address: bc1q98q4yn0ea9cvuthgam7lt250ct3vln65zytfae
ETH Address: 0x0Ce35988F785524697AE11D28cdd4A1b97123f11
