README.md

---
language:
- en
pipeline_tag: text-generation
tags:
- text-generation-inference
- instruct
license: apache-2.0
---
<h1 style="text-align: center">Eros-7B-Test (WIP Name)</h1>
<h2 style="text-align: center">Experimental Roleplay Finetine</h2>

## Model Details
**This is considered an unofficial model**. 

An experimental model that uses a new version of PIPPA dataset as the primary base. This PIPPA dataset is the original one we have uploaded that has been refined, augmented and trimmed down for proper model training. 
The model is a finetune on the Mistral-7B base with 22K token examples. Eros-7B is primarily designed for ChatRP and with some capabilities to do story generations too. It is trained on the ChatML format.  

Due to it being an experimental model, there are some quirks...

- Rare occasion to misspell words
- Rare occasion to have random formatting artifact at the end of generations
- Tendencies to use the same phrase when generating (e.g. *she was always smiling* variants persisting in multi-turn conversations)
- Not very smart but highly creative due to a lack of logic/reasoning dataset

While this model is not good enough to be deemed as an official release model under the PygmalionAI name, I feel like it is a good stepping point to give this to the public under this account. Any feedback is appreciated. The above mentioned issues will be fixed in the next training attempt of models.

## Prompting Details
**This is under the assumption this model is used with [SillyTavern](https://github.com/SillyTavern/SillyTavern), please note it may not cover other existing application use cases.**

Use the ChatML Instruct Settings
<img src="https://files.catbox.moe/6318gp.png" alt="sillytavernsettings" width="350" height="500">
Use these settings for consistent generations
<img src="https://files.catbox.moe/ayos28.png" alt="sillytavernsettings" width="350" height="500">

**Note**: Temperature, and Min P values can be adjusted to greater or lower values depending on generation preferences. 

## Limitations and biases
The intended use-case for this model is fictional writing for entertainment purposes. Any other sort of usage is out of scope.

As such, it was **not** fine-tuned to be safe and harmless: the base model _and_ this fine-tune have been trained on data known to contain profanity and texts that are lewd or otherwise offensive. It may produce socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive. Outputs might often be factually wrong or misleading.