Unknown training token count for Mistral 7B

Ascertain the number of training tokens used to train Mistral 7B to enable precise comparisons with Llama-2 and Solar and to evaluate the role of training data volume in preference learning and brittleness.

Background

The discussion relates model brittleness and value-based preference emergence to training volume and attention architecture. Llama-2 reportedly used 2 trillion tokens and Solar built on Mistral weights, but the exact training token count for Mistral remains unspecified, limiting causal inference about token volume effects.