Introducing Quasar 1.0 – our strongest model

Introducing Tavik-1.3

Tavik-1, trained further for step-by-step math and logic – public on Hugging Face.

Tavik-1.3

Tavik-1.3 is a new version of Tavik-1. We trained it further on step-by-step arithmetic, logic puzzles, its original conversations and who it is. It keeps the same 20.4 million parameters.

Highlights

  • 25.2% on our internal arithmetic evaluation (Tavik-1: 0%)
  • 52 of 100 logic tasks solved (Tavik-1: none)
  • 10.0% on Question answering · Basis in Convergent 3.0 (Tavik-1: 6.0%)
  • 11.1% on Math · Bonus in Convergent 3.0 (Tavik-1: 7.6%)

Good to know

On Math · Basis in Convergent 3.0 it scores 3.5%, lower than Tavik-1 (6.3%). It still cannot write code, and longer creative texts such as poems can repeat themselves.

Tavik-1.3 is public on Hugging Face.

Links