metadata

license: apache-2.0
datasets:
  - PocketDoc/Retro-YahooAnswers
language:
  - en
pipeline_tag: question-answering
base_model: mistralai/Mistral-7B-v0.1

Description

Do you miss the vibes of the early 2000s? Yearn for the nostalgia of internet religious arguments? Then this model is for you!

This was trained on a scrape of Yahoo! Answers from 2007 and received no filtering save for basic sanity checks.

This is not intended for serious use but I think it's charming in a way.

Prompt format:

Pygmalion / 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 quick and dirty training details:

Sequence length: 2048
Training time: 32 hours
Hardware: 1x RTX 4080
Training type: QLoRA
PEFT R/A: 32/32