license: cc-by-2.0
language:
- en
tags:
- finance
- legal
- biology
- art
Behold, one of the first fine-tunes of Mistral's 7B 0.2 Base model. SatoshiN is trained on 4 epochs 2e-4 learning rate (cosine) of a diverse custom data-set, combined with a polishing round of that same data-set at a 1e-4 linear learning rate. It's a nice assistant that isn't afraid to ask questions, and gather additional information before providing a response to user prompts.
I have found varying success using instruction-formats such as Alpaca, ChatML and Mistral. The custom training was performed on raw-text with the idea that it might acquire better generalization skills.
Total model-size has increased from 7.24B to 7.35B after merging a .5GB LoRa via PEFT.
SatoshiN | Base-Model
Wikitext Perplexity: 6.27 | 5.4
**Similar to SOTA, this model runs a bit hot, try using lower temperatures below .5 if experiencing any nonsense)