Description:
This is a multipurpose chat / chat instruct hybrid model in the same vein as the Pygmalion team's Metharme. It uses a curated pile of training data that has been normalized into a consistent training format. It has been trained on a wide array of one shot instructions, multi round instructions, and role playing scenarios.
The training parameters were suboptimal for the most recent run and I decided to stop after 2 epochs as 3 likely would have overtrained it. I plan on iterating the model and improving it further when I have access to more funds to do so.
Prompt format:
Metharme
The prompt should start with the cursor on the same line directly after "<|model|>" with no space. The following are all valid formats and can be extended to as many rounds as desired.
<|system|>system message here<|user|>user message here<|model|>
<|system|>system message here<|user|>user message here<|model|>model message<|user|>user message here<|model|>
<|system|>system message here<|model|>
<|system|>system message here<|model|>model message<|user|>user message here<|model|>
Some example prompts:
<|system|>The following is a transcript between a helpful assistant and a user.<|user|>Why is the sky blue?<|model|>
<|system|>You are a Virtual Story Generator. You take the user's input and create an excellent and captivating story that goes in that direction. Use an abundance of sensory descriptions and eloquent prose.<|user|>Alpha Centauri has fallen, to the bears. This is a point of view tale about a soldier on the ground.<|model|>
<|system|>You are a professional editor with decades of experience, help the user with any task they have for you.<|user|>Can you rewrite this to flow better? "I knew I probably shouldnt have done that but oh well"<|model|>
More will be added at a later date.
Perplexity Benchmarks:
- TBA
Training information:
- GPTQ 4 bit LoRA
- 2 Epochs
- 64 / 32 R / A
- 2048 Cutoff
- 42 hours on 1x RTX 4090
Data used in training:
- TBA
Models used:
For training: https://huggingface.co/PocketDoc/llama-30b-gptq-4bit-128g
For merging:
https://huggingface.co/PocketDoc/Dans-PersonalityEngine-30b-LoRA
and
https://huggingface.co/huggyllama/llama-30b
Disclaimer:
It has not been aligned and no warranty is given for the quality or safety of its outputs.