Voxtral-TCR1-4b

This is a Voxtral model with removed voice module and finetuned to give it a custom task-concept reasoning pattern:

Question: <question>

Sub-tasks:
    1.
    2.
    3.

Key concepts:
    -
    -
    -

Use ChatML formatting, force thinking mode in your favourite front-end (prefill with <think> token).

Temps in range 0.6-0.8 seem to work reasonably well.

This is an experiment to see if thinking/reasoning could be bootstrapped from 0 without any reasoning datasets whatsoever. The answer is yes.

This model is trained on purely artificial data of non-reasoning models.