reeducator / bluemoonrp-13b

General

Bluemoon roleplay finetune of LLaMA 13B (2 roleplayers only).

Models

Two models are provided, labeled (1) 4k-epoch6 and (2) epoch3 (other branch). In case of the (1), the training is extended over more epochs to reduce the high training loss observed in (2). This release also tests a longer 4k context token size achieved with AliBi.

GGML 4-bit for llama.cpp

1. ggml-bluemoonrp-13b-4k-epoch6-q5_0.bin
2. ggml-bluemoonrp-13b-epoch3-q5_0.bin

GPTQ 4-bit CUDA:

1. bluemoonrp-13b-4k-epoch6-4bit-128g.safetensors
2. bluemoonrp-13b-epoch3-4bit-128g.safetensors

Remarks

This model has been trained using the following prompt (Vicuna 1.1 format):

A transcript of a roleplay between two players, LEAD and ASSOCIATE. LEAD sets up a scenario and the characters, from which ASSOCIATE then assumes a character role and continues the story for that role in response to description given by LEAD. The story and characters are developed by exchange of detailed event descriptions and character dialogs, successively given by both LEAD and ASSOCIATE.
LEAD: [role1 message]
ASSOCIATE: [role2 message]</s>