|
|
|
|
@@ -4,7 +4,7 @@ Allows you to play roleplay scenarios with large language models.
|
|
|
|
|
|
|
|
|
|
It does not run any large language models itself but relies on existing APIs. Currently supports **text-generation-webui** and **openai**.
|
|
|
|
|
|
|
|
|
|
This means you need to either have an openai api key or know how to setup [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) (locally or remotely via gpu renting.)
|
|
|
|
|
This means you need to either have an openai api key or know how to setup [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) (locally or remotely via gpu renting. `--api` flag needs to be set)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
@@ -18,15 +18,17 @@ This means you need to either have an openai api key or know how to setup [oobab
|
|
|
|
|
- summarization
|
|
|
|
|
- director
|
|
|
|
|
- creative
|
|
|
|
|
- multi-client (agents can be connected to separate LLMs)
|
|
|
|
|
- long term memory (very experimental at this point)
|
|
|
|
|
- multi-client (agents can be connected to separate APIs)
|
|
|
|
|
- long term memory (experimental)
|
|
|
|
|
- chromadb integration
|
|
|
|
|
- passage of time
|
|
|
|
|
- narrative world state
|
|
|
|
|
- narrative tools
|
|
|
|
|
- creative tools
|
|
|
|
|
- AI backed character creation with template support (jinja2)
|
|
|
|
|
- AI backed scenario creation
|
|
|
|
|
- runpod integration
|
|
|
|
|
- overridable templates foe all LLM prompts. (jinja2)
|
|
|
|
|
- overridable templates for all prompts. (jinja2)
|
|
|
|
|
|
|
|
|
|
## Planned features
|
|
|
|
|
|
|
|
|
|
@@ -34,20 +36,27 @@ Kinda making it up as i go along, but i want to lean more into gameplay through
|
|
|
|
|
|
|
|
|
|
In no particular order:
|
|
|
|
|
|
|
|
|
|
- Gameplay loop governed by AI
|
|
|
|
|
- Extension support
|
|
|
|
|
- modular agents and clients
|
|
|
|
|
- Improved world state
|
|
|
|
|
- Dynamic player choice generation
|
|
|
|
|
- Better creative tools
|
|
|
|
|
- node based scenario / character creation
|
|
|
|
|
- Improved long term memory (base is there, but its very rough at the moment)
|
|
|
|
|
- Improved and consistent long term memory
|
|
|
|
|
- Improved director agent
|
|
|
|
|
- Right now this doesn't really work well on anything but GPT-4 (and even there it's debatable). It tends to steer the story in a way that introduces pacing issues. It needs a model that is creative but also reasons really well i think.
|
|
|
|
|
- Gameplay loop governed by AI
|
|
|
|
|
- objectives
|
|
|
|
|
- quests
|
|
|
|
|
- win / lose conditions
|
|
|
|
|
- Automatic1111 client
|
|
|
|
|
|
|
|
|
|
# Quickstart
|
|
|
|
|
|
|
|
|
|
## Installation
|
|
|
|
|
|
|
|
|
|
Post [here](https://github.com/final-wombat/talemate/issues/17) if you run into problems during installation.
|
|
|
|
|
|
|
|
|
|
### Windows
|
|
|
|
|
|
|
|
|
|
1. Download and install Python 3.10 or higher from the [official Python website](https://www.python.org/downloads/windows/).
|
|
|
|
|
@@ -64,7 +73,7 @@ In no particular order:
|
|
|
|
|
1. `git clone git@github.com:final-wombat/talemate`
|
|
|
|
|
1. `cd talemate`
|
|
|
|
|
1. `source install.sh`
|
|
|
|
|
1. Start the backend: `python src/talemate/server/run.py runserver --host 0.0.0.0 --port 5001`.
|
|
|
|
|
1. Start the backend: `python src/talemate/server/run.py runserver --host 0.0.0.0 --port 5050`.
|
|
|
|
|
1. Open a new terminal, navigate to the `talemate_frontend` directory, and start the frontend server by running `npm run serve`.
|
|
|
|
|
|
|
|
|
|
## Configuration
|
|
|
|
|
@@ -108,7 +117,7 @@ Will be updated as i test more models and over time.
|
|
|
|
|
| [Synthia v1.2 34B](https://huggingface.co/TheBloke/Synthia-34B-v1.2-GPTQ) | 34B model | Cannot be run at full context together with chromadb instructor models on a single 4090. But a great choice if you're running chromadb with the default embeddings (or on cpu). |
|
|
|
|
|
| [Xwin-LM-70B](https://huggingface.co/TheBloke/Xwin-LM-70B-V0.1-GPTQ) | 70B model | Great choice if you have the hardware to run it (or can rent it). |
|
|
|
|
|
| [Synthia v1.2 70B](https://huggingface.co/TheBloke/Synthia-70B-v1.2-GPTQ) | 70B model | Great choice if you have the hardware to run it (or can rent it). |
|
|
|
|
|
| [GPT-4](https://platform.openai.com/) | Remote | Still the best for consistency and reasoning, but is heavily censored. While talemate will send a general "decensor" system prompt, depending on the type of content you want to roleplay, there is a chance your key will be banned. **If you do use this make sure to monitor your api usage, talemate tends to send a lot more requests than other roleplaying applications.** |
|
|
|
|
|
| [GPT-4](https://platform.openai.com/) | Remote | Still the best for consistency and reasoning, but is heavily censored. Talemate will send a general "decensor" system prompt, ymmv. **If you do use this make sure to monitor your api usage, talemate tends to send a lot more requests than other roleplaying applications.** |
|
|
|
|
|
| [GPT-3.5-turbo](https://platform.openai.com/) | Remote | It's really inconsistent with JSON responses, plus its probably still just as heavily censored as GPT-4. If you want to run it i'd suggest running it for the conversation agent, and use GPT-4 for the other agents. **If you do use this make sure to monitor your api usage, talemate tends to send a lot more requests than other roleplaying applications.** |
|
|
|
|
|
|
|
|
|
|
I have not tested with Llama 1 models in a while, Lazarus was really good at roleplay, but started failing on JSON requirements.
|
|
|
|
|
|
