README.md

---
license: apache-2.0
datasets:
- N8Programs/CreativeGPT
language:
- en
pipeline_tag: text-generation
---

![image/png](https://cdn-uploads.huggingface.co/production/uploads/647e98971a1fcad2fdc55e61/JeTZMxno0ANlYLGLzvX2X.png)

# Model Card for Coxcomb

A creative writing model, using the superb [senseable/WestLake-7B-v2](https://huggingface.co/senseable/WestLake-7B-v2) as a base, finetuned on GPT-4 outputs to a diverse variety of prompts. It in no way competes with GPT-4 - it's quality of writing is below it, and it is primarily meant to be run in offline, local environments.
On creative writing benchmarks, it is consistently ranked higher than most other models - [it scores 72.37](https://eqbench.com/creative_writing.html), beating goliath-120b, yi chat, and mistral-large.
It is designed for **single-shot interactions**. You ask it to write a story, and it does. It is NOT designed for chat purposes, roleplay, or follow-up questions.

## Model Details

Trained w/ a 40M parameter lora on [N8Programs/CreativeGPT](https://huggingface.co/datasets/N8Programs/CreativeGPT) for 3 epochs. Overfit slightly (for much better benchmark results).


### Model Description

- **Developed by:** N8Programs
- **Model type:** Mistral
- **Language(s) (NLP):** English
- **License:** Apache 2.0
- **Finetuned from model:** [senseable/WestLake-7B-v2](https://huggingface.co/senseable/WestLake-7B-v2)

## Uses

Bot trained on NSFW (sexual or violent) content but will generate it when asked - it has not been trained with refusals. If you wish to ADD refusal behavior in, further tuning or filtering will be neccessary.

### Direct Use

GGUFs available at [Coxcomb-GGUF](https://huggingface.co/N8Programs/Coxcomb-GGUF)
Should work with transformers (not officially tested).

## Bias, Risks, and Limitations

Tends to generate stories with happy, trite endings. Most LLMs do this. It's very hard to get them not to.

## Training Details

Trained on a single M3 Max in roughly 12 hours.