BestERP

Psyfighter v2 13B

overall
0.7
(28)
nsfw
0.8
(17)
role playing
0.8
(16)
story writing
0.8
(14)

Overview

Psyfighter v2 13B is a jack of all trades and a master of some. The goal was to improve prose and logic capabilities over Tiefighter 13B. To that end, the model includes CatV1 and LimaRP.

It shines the most in story mode, where, according to the authors, it "spits out pure gold." Try giving the model a URL (even a fake one) of the form www.books.com/author/genre/title and it will try to match its output to the author/genre/title.

Please share your experiences with this model in a review here, or directly with the model creators Jeb Carter and TwistedShadows from the KoboldAI community at their Discord: https://koboldai.org/discord

Usage

This is a summary of the model's HuggingFace page.

This model is meant to be creative, If you let it improvise you get better results than if you drown it in details.

Story Writing

The model can do regular story completion task, where you place your existing story into the prompt, and let the model continue. You can steer the model through instructions or author notes.

Generate story with given topic or theme

You can use the Alpaca format, for example:

### Instruction: 

Write a novel about X, use chapters and dialogue

### Response: 

The best way to control the output style and format (1st person vs 3rd person, how much dialog there is, etc.) is to provide the model with the start of the story in your desired style and format.

Chatbots and role-play

With Psyfighter V2, less is more, because the model is great at improvising. Instead of drowning the model in paragraphs of context, try to keep it simple and let the model help you figure out the character.

Try putting this in memory in regular chat mode:

### Instruction: 

Generate a conversation between Alice and Jeb where they discuss language models.
In this conversation Henk is excited to teach Alice about Psyfighter. 

### Response: 

Using chat mode or Alpaca is recommended, but other instruct formats might work as well, since the model is a merge of many other models.

Adventure Games

The model contains a LoRA trained on KoboldAI's Skein adventure dataset. The best way to do adventures, is to provide an introduction to your world and your objective, and use the > prefix for user commands (a.k.a. KoboldAI's adventure mode).

If the model does not pick-up on the style, try correcting the output and continuing from there.

Instruct Prompting

This model was designed for fiction (AI characters, role-play, story-writing, etc.), not for general information seeking tasks. As such, it might be prone to hallucinations, and you may see some of the fiction data leaking into the output.

Submit your review

2.5 Not good

Pick which version of the model you base your review on.

Reviews

2.0
quant: gguf q6

Creative but frequently gets stuck on certain details, mixes up details between characters and frequently misgenders.

2.0
quant: gguf q5

Good for storytelling but lacks at RP due to purple prose and too much verbosity.

3.5
quant: gguf q5

This is my first ever model coming out of agnaistic, GPT 3.5, and Kobold. There are things good with this one, I've tried many characters on it, and on average it stays in character. One thing I found hard to keep details is time, dinner can suddenly turn into breakfast in 2 prompts. I don't mind the verbosity, sometimes it could be a better way to explain what's currently happening. The last one is that at least 4 hate/fear characters I have tried could turn to 'I love you' in less than 10 prompts.

3.0
quant: gguf q5

I’ve downloaded this a few times, and frankly I just never quite found it up to begin at with other models. Not sure if it was my instruction set or what, but this model just was a bit meh. I found the original a bit better overall, and just couldn’t take this thing and it’s hallucinations. There’s better models out there I think that represent more wholly what a model of this size is capable of.

4.0
quant: gguf q5

Imaginative but no nonsense, stays in the line with story.

1.5
full precision

this is good and all, till a certain point.

0.5

This is absolutely garbage. The only thing that its good for is maybe basic conversation. Otherwise, I don't recommend this model.