Helpful Links
If you're new to Pygmalion, take a look at the Official Documentation Page first. This will answer most of your basic questions.
To access Pygmalion, choose one of the following Colab links and follow the launch instructions to get started: TextGen WebUI - simple CAI-like interface TavernAI - feature-rich atmospheric interface Local installation guides are below.
The following is a compendium of links, guides, and tools related to Pygmalion and similar Large Language Models (LLMs). The field of LLMs is ever-changing. Some of the links below may already be outdated, but we will try our best to keep this article up to date.
Pygmalion Links
Pygmalion Models
- List of Pygmalion models
- Pygmalion 13B | A conversational LLaMA fine-tune.
- XOR weight conversion | These XOR conversions are necessary since the original model weights can't be used as-is, due to licensing issues by Meta (creators of LLaMA).
- Pygmalion 13B 4bit | 4bit quantized models will use far less GPU resources, at a small precision cost.
- Pygmalion 13B GGML | GGML quantization allows models to run entirely on CPU, no $2000 GPU necessary.
- Metharme 13B
- Pygmalion 7B | A conversational LLaMA fine-tune.
- Metharme 7B | An instruction-tuned LLaMA biased towards fiction writing and conversation.
- Pygmalion 6B | Based on EleutherAI's GPT-J 6B
- Pygmalion 2.7B | Based on EleutherAI's GPT-Neo 2.7B
- Pygmalion 1.3B | Based on EleutherAI's Pythia 1.3B
- Pygmalion 350M | Based on Facebook's OPT 350M
- Documentation page for 13B and 7B
- Run GGML quantized models | You can run both LLaMA and GPT-J versions with this. (Moved to this link from the currently unsupported link).
- GGML quantized 7B and 6B models
- GPTQ 4bit quantized 6B models
User Interfaces
The old GradioUI colab is closed. Use one of the alternative UIs below.
Google has cracked down on PygmalionAI in Colab. If the Colab uses the string PygmalionAI
in its code, it won't work. You can still use Colab with other language models, and you can still run PygmalionAI locally. The main Colabs now use different versions of PygmalionAI, which aren't banned yet.
Developers have been working on an official UI, nicknamed Galatea. Also known as "The Website". It will eventually be hosted here, once they acquire the necessary hardware. Have patience. Remember: they're doing this for free, in their own spare time.
In the meantime, you can enjoy Pygmalion using one of the following interfaces:
TextGen WebUI
A combined front-end and back-end for running models like LLaMA, GPT-J, Pythia, etc.
Link | Notes |
---|---|
Colab | Simple CAI-like interface with a few optional parameters that you can check before running. Works well on mobile. |
Local Installers | One-click installer. For Windows, Linux, and MacOS systems. |
Installation guide | By Oobabooga |
Installation guide | By PygmalionAI |
Github | |
Wiki | Lots of useful info on things like DeepSpeed, FlexGen, LLaMA, RWKV, LoRAs, and more. |
Subreddit | |
---- | |
### TavernAI | |
An atmospheric interface with lots of customization options, that uses a card-based character file system. Note that TavernAI is only a front-end; it needs to connect with a back-end such as Kobold (see below) or Oobabooga (see above). | |
Link | Notes |
------ | ------ |
Alpin's Colab | This is a fork of the original TavernAI Colab, since the original Colab no longer supports Pygmalion models. |
Mobile Colab | Another fork, with a few changes for mobile users, such as better UI scaling. Runs on an older Tavern version. |
SillyTavern | A Tavern fork under active development that adds many more features over the original Tavern. To use SillyTavern, see the installation guide here. And for Android users, here is an installation guide using Termux (a Linux emulator). Note: For security reasons, the SillyTavern colab notebook will only host the Extras Extensions API, not the actual Web UI; to use the Web UI, you need to install the SillyTavern files to your computer first. |
TavernAI Github | |
SillyTavern Subreddit | |
Guide for mobile users | |
---- | |
### KoboldAI | |
A back-end for running a wide range of open language models. | |
Link | Notes |
------ | ------ |
TPU Colab | Runs larger models. |
GPU Colab | Runs smaller models. |
KoboldAI Horde Client | More info on AI hordes. |
Kaggle | Alternative to Colab, but can be very janky at times. You can use these SMS tools if you don't want to give out your phone number. |
koboldcpp | C++ implementation |
Installation guide (GPU) | By PygmalionAI |
Installation guide (4bit) | For GPUs with less than 10GB of VRAM |
KoboldAI Github | United fork of the original KoboldAI page and several other forks. Generally more updated. |
Subreddit | |
---- | |
### Other UIs | |
Link | Notes |
------ | ------ |
miku.gg | Open-source UI to run chatbots. Allows visual and text-to-speech (TTS). (How to make emotion sprites for Miku bots) |
RenAI Chat | Virtual Novel interface to chat with AI models. |
AgnAIstic | UI to run chatbots anonymously, no login required. (Github link) (Local installation guide, NSFW) |
Poe | Made by Quora. A free way to talk to chatbots. Includes ChatGPT/GPT-4 (by OpenAI), and Claude+/Claude-instant (by AnthropicAI). |
MonikA.I. | "This project aims to add new AI based features to Monika After Story mod with the submod API." |
OpenCharacters | Create and share ChatGPT/AI characters. |
Laika's LiteVN UI | Highly customizable visual novel-like UI. |
Web UI with LTM | Experimental Web UI with Long Term Memory (LTM), which helps the bot remember long-term conversations. |
Open Playground | An LLM playground you can run on your laptop. Github page |
VRoid Hub | Alternative to WebUI - Unity prototype with VRoid 3D |
Your AI Friend | Simple UI, similar to Replika |
Faraday.dev | Run open-source LLMs on your computer. Completely offline. Zero configuration. |
VenusAI | Free uncensored frontend for OAI and Kobold. |
RisuAI | Tavern-like frontend, feature-rich and has plugins. Run in web or with installer. Github page |
---- | |
## Guides / Tips | |
### General Guides | |
Link | Notes |
------ | ------ |
Alpin's Pygmalon Guide | Very thorough guide for installing and running Pygmalion on all types of machines and systems. Spiritual successor to the old rentry guide. |
Official documentation page | Contains guides and info on Pygmalion. Currently hosted on Github, but will be moved to the official website when it releases. You are encouraged to contribute to the documentation page here. |
Local Installation Guide (CPU) | You can run Pygmalion on CPU only, with only 4-8GB of RAM (not VRAM) required, thanks to GGML. No $2000 GPU necessary. |
"Okay, so what the heck is a chatbot?" | Very simple and helpful guide to getting a chatbot up and running on Tavern. |
CAI chatlog dumping | Upload your CAI dumps here to help train Pygmalion. |
Quick Guide to Pygmalion for Colab | Recommended for those new to Colab. |
Visual guide for Colab | For people who learn better with pictures. |
Soft Prompts | Read more about soft prompts here |
General guide to open-source chatbots | General chatbot guide with very helpful tips and explanations. Also expands to language models outside of Pygmalion. |
### Character Guides | |
Link | Notes |
------ | ------ |
Pygmalion Tips | Tips and information for getting better responses and personalities out of your characters. |
Pygmalion character creation + writing tips | Helpful guide on character creation. |
How to make a character | Another helpful character creation guide. (Trim down those wall-of-text character descriptions!) |
Generating better responses | General usage, character creation guide, memory management, rumor-busting. |
Importing your CAI character to Pygmalion | |
Continuing your CAI RP using Pygmalion | |
---- | |
##Character Repositories | |
Link | Notes |
------ | ------ |
BotPrompts | View, submit, and request JSON files for Pygmalion |
Tavern card booru | Share and download Tavern cards (NSFW) |
Character Hub | Share and download Tavern cards (SFW; has a NSFW toggle) |
/aids/ Prompts | Story prompts for KoboldAI bots |
---- | |
## Tools / Scripts | |
Link | Notes |
------ | ------ |
Pygmalion creator tool | Enter your CAI character info to create Pygmalion-compatible JSON files |
AI Character Editor | Create and convert to/from CAI, Pygmalion, KoboldAI, and TavernAI formats. Supports JSON and image files. |
TAIConvert | Convert TavernAI cards to JSON files to use in other UIs |
CAI to Pygmalion | Convert CAI chatlogs to be compatible with Pygmalion |
CAI log/definition downloader | Allows you to download all of your CAI bot's chat history, as well as their definitions (if you have access to them). |
CAI tools | A Chrome extension made for CAI data scraping |
More helpful CAI tools and scripts | HYW, fix tildes, darken italics, and more. |
W++ formatter | See this example image for a general idea of what to do. W++ is supposedly a more machine-readable prompt format for Pygmalion. You can see some example W++ characters here. |
W++ to Boostyle converter | Boostyle is a character prompt format that is supposedly as effective as W++. More on Boostyle. |
---- | |
## Other models and chatbots | |
- Awesome Marketing Datascience | A curated list of useful LLMs / analytics / data science resources |
- Open LLM Leaderboard | The HuggingFace Open LLM Leaderboard tracks, ranks and evaluates LLMs and chatbots as they are released. |
- There's An AI For That | Massive list of user-submitted AIs from a wide variety of categories, including conversational, image gen, audio, and a lot more. The site itself was written by ChatGPT. |
- OpenAssistant | Open source alternative to ChatGPT, in development by LAION. You can contribute here. Subreddit link |
- Awesome Totally Open Chatgpt | A list of open-source alternatives to ChatGPT |
- Pickaxe Project | Build, manage, and deploy custom GPT-4 apps |
- MPT-7B | Model trained on 1 trillion tokens, by MosaicML |
- MPT-7B Storywriter | Model with 65K-token context length |
- PoemiseAI | Free GPT-4 content generator |
- ChatGLM | A bilingual (English/Chinese) open-source model based on the General Language Model (GLM) framework. (Online demo). |
- Cerebras GPT | A family of open, compute-efficient LLMs, from 111M to 13B parameters. |
- StableLM | An open source ChatGPT alternative by Stability AI |
- Yandere AI Simulator | Escape Room-style game with a yandere girlfriend powered by ChatGPT. Requires an OpenAI API key to play. |
- ColossalChat | Open-source alternative to ChatGPT by ColossalAI, based on LLaMA, fine-tuned, with RLHF pipeline. See this article. |
- EmbedBase | Open-source SDK and API to easily connect data to ChatGPT (Github link) |
- Auto-GPT | An experimental open-source program that uses GPT-4 to write and debug its own code. See this twitter thread. |
- BLOOM | Open-access multilingual LLM |
- YaLM-100B | Pretrained language model with 100B parameters, by Yandex |
- PaLM RLHF Pytorch | "Basically ChatGPT but with PaLM" |
- LaMDA RLHF Pytorch | Open-source pre-training implementation of Google's LaMDA in PyTorch. |
###LLaMA | |
- LLaMA | LLM created by Meta. The model was inadvertently leaked a few days after its announcement. Needless to say, this has been a massive boon for the LLM community. So much so, I dedicated several sections to it. |
- Subreddit | |
- Simplified LLaMA installation guide | |
- The Extremely Simplified First-time LLaMA guide | |
- LLaMA common questions | |
- Guide to LLaMA roleplay | |
- Using ChatGPT plugins with LLaMA | |
- LLaMA C++ implementation | |
- LLaMa Rust implementation | |
- Llama Hub | A library of custom data loaders/readers for LLMs. (Github link) |
####LLaMA models | |
- HuggingFace | List of LLaMa models |
- Koala | A chatbot trained by fine-tuning LLaMA on dialogue data gathered from the web. |
- WizardLM | "An Instruction-following LLM Using Evol-Instruct" (Uncensored version) |
- Vicuna | "An open-source chatbot trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT." (Github) (Demo) |
- Wizard Vicuna 13B Uncensored 4Bit | "GPTQ format quantized 4bit models of Eric Hartford's 'uncensored' training of Wizard-Vicuna 13B." |
- Manticore 13B | An updated version of Wizard Mega 13B. A fine-tuned Llama 13B model available in GPTQ and GGML quantizations. See a GGML-powered demo here |
- GPT4All | "A chatbot trained on a massive collection of clean assistant data including code, stories and dialogue." (Not related to GPT-4) |
- OpenBuddy | Multilingual LLaMA-based chatbot |
- HuggingChat | LLaMA-based chatbot + interface |
- GPT4All-13B Snoozy | GGML format |
####Alpaca | |
- Stanford Alpaca | An instruction-following LLaMA model |
- Alpaca-LoRA | Reproducing the results of Stanford Alpaca on consumer hardware with the use of Low-Rank Adaptation (LoRA) |
- Alpaca Colab link | This notebook contains minimal code for running Alpaca-LoRA for demonstration purposes. |
- HuggingFace link | List of Alpaca models |
- Alpaca C++ implementation | |
- Test out Alpaca | |
- Dalai | Easy local installation |
- Code Alpaca Demo | |
- Video guide by Aitrepreneur | |
---- | |
##Additional resources | |
- FlexGen | Run large models on consumer-grade GPUs |
- RLHF | Reinforcement Learning with Human Feedback, a method to help LLMs learn and self-improve. |
- CoH | Chain of Hindsight, another self-improving method for LLMs. |
- Guidance | Prompting language by Microsoft for "guiding" samplings from open LLMs |
- Simon Willison's Weblog | News on LLMs and machine learning |
- Machine Learning subreddit | |
- LocalLLM Wiki | |
- Together | |
- RedPajama | "RedPajama is a project to create a set of leading, fully open-source models. " |
- Local LLM links | |
- Local LLM links | |
- LLM papers | |
---- | |
Keep your Smile. Keep on Smiling. |