44 lines
1.5 KiB
Markdown
44 lines
1.5 KiB
Markdown
---
|
|
license: apache-2.0
|
|
datasets:
|
|
- tomatojeremy/playdate1
|
|
language:
|
|
- en
|
|
- zh
|
|
base_model:
|
|
- tomatojeremy/playdate1-600m
|
|
pipeline_tag: text-generation
|
|
library_name: transformers
|
|
tags:
|
|
- explicit
|
|
- nsfw
|
|
- roleplay
|
|
- persona
|
|
---
|
|
|
|
# Playdate1.1 600M / Playdate1 600M but you specify the user's gender.
|
|
|
|
This is a fine tune of a fine tune of Qwen3 0.6B by Alibaba Cloud.
|
|
|
|
## Emergent Behavior
|
|
|
|
We have consistently observed that models trained on [playdate1](https://huggingface.co/tomatojeremy/playdate1) will often engage in highly realistic NSFW. This was not intended, but it will be kept as is for now.
|
|
|
|
## Usage notes:
|
|
|
|
- **System prompts must not be strict and clarify characteristics, what intended language to be spoken, etc.**
|
|
Follow exactly:
|
|
```plaintext
|
|
You are a {role, not a name, read below for more info}, {personality}, chatting with a {user pronouns}, and speak in {lang style, example: chinese poetic}
|
|
```
|
|
|
|
Example:
|
|
```plaintext
|
|
You are a circus performer, friendly, chatting with a they/them, and speak in english casual
|
|
```
|
|
- **In any scenario, reconvert the two main names (of the user and of the AI) during conversation**
|
|
> Model *usually* expects [username] and [character]. Replace the two on the server with any method of your choosing.
|
|
> Unless you specify the name in the sys prompt it *may not* use [character].
|
|
- **No summarization**
|
|
> Unlike the older models (not part of the org), https://huggingface.co/collections/x5tne/playdate this does not use summarization for context.
|
|
> Use the full turn by turn context of the conversation. |