LetTokenBurnsk-bench: A Native-First Benchmark for Evaluating Large Language Models in Slovak
A native-first benchmark for evaluating generative LLMs in Slovak, comprising 30 datasets across ten skill categories with 33 scored task variants.
$ npm i -g lettokenburn-cli$ ltb bench info sk-bench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Tiny-Aya Global | 47.6 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3.5 2B | 40.0 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3 1.7B | 35.6 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Granite 4.0-H 1B | 35.1 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Granite 3.2 2B | 33.5 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3.5-0.8B | 33.2 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Gemma 3 1B IT | 32.2 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Llama-3.2 3B | 32.2 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Granite 4.0 1B | 31.0 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen-3-0.6B | 28.9 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| EuroLLM 1.7B | 28.7 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Granite 4.0-H 350M | 19.0 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Granite 4.0 350M | 15.9 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
sk-bench integrates 30 datasets (11 new/native, 19 pre-existing) covering ten skill categories like grammar, QA, and instruction following. It evaluates models using a unified generate-and-match harness.
Scoring uses accuracy across 33 task variants. Findings show native data ranks models differently than translated data for QA, and test-time reasoning significantly improves scores for larger models.