Description
This is an instruction following model (based on Vikhr Base) optimized for Russian language. It was trained using kolibrify via curriculum learning on a multitude of instruction datasets.
The model uses ChatML template. It was trained to be sensitive to the system prompt, experiment with it.
Primary uses: information retrieval, classification, semantic analysis, translation (en-ru), etc. Currently in pre-alpha.
Instruction following evals
The model was tested using the following benchmarks:
| Eval name | Strict Value | Loose Value |
|---|---|---|
| Avg. | 37.82 | 39.24 |
| ifeval-prompt-level | 33.08 | 35.12 |
| ifeval-instruction-level | 44.48 | 46.40 |
| ru-ifeval-prompt-level | 31.79 | 32.53 |
| ru-ifeval-instruction-level | 41.96 | 42.92 |