🇨🇿 CzechBench Leaderboard
The goal of the CzechBench project is to provide a comprehensive and practical benchmark for evaluating Czech language models. Our evaluation suite currently consists of 15 individual tasks, leveraging pre-existing Czech datasets together with new machine translations of popular LLM benchmarks, including ARC, GSM8K, MMLU, and TruthfulQA. This work is brought to you by CIIRC CTU and VSB Ostrava.
Key Features and Benefits:
- Tailored for the Czech Language: CzechBench includes both original Czech datasets and adapted versions of international datasets, ensuring relevant evaluation of model performance in the Czech context.
- Wide Range of Tasks: It contains 15 different tasks that cover various aspects of language understanding and text generation, enabling a comprehensive assessment of the model's capabilities.
- Bilingual performance analysis: CzechBench also offers a parallel collection of 9 English tasks corresponding to the Czech versions included in the main suite. This allows for direct comparison of model performance across both languages with equivalent conditions in terms of prompt formulation and few-shot example selection.
- Universal model support: The universal text-to-text evaluation approach adopted in CzechBench allows for direct comparison of models with varying levels of internal access, including commercial APIs.
- Ease of Use: The benchmark is built upon a commonly used evaluation framework with wide support for state-of-the-art models and inference acceleration tools.
- Empowering decisions: Whether you are a business looking for the best LLM solution to base your application on, or a research team trying to maximize the capabilities of the models they are developing, CzechBench will help you gain insights into particular strengths and weeknesses of individual models and better focus on key areas for optimization.
Below, you can find the up-to-date loaderboard of models evaluated on CzechBench. For more information on the included benchmarks and instructions on evaluating your own models, please visit the "About" section below.
The values shown in the leaderboard table represent the accuracy metric in percentage.
Model | Precision | Aggregate Score | Grammar (Avg.) | Knowledge (Avg.) | Reasoning (Avg.) | Math (Avg.) | Classification (Avg.) | AGREE | ANLI | ARC-Challenge | ARC-Easy | Belebele | CTKFacts | Czech News | Facebook Comments | GSM8K | Klokanek | Mall Reviews | MMLU | SQAD | Subjectivity | TruthfulQA | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
bfloat16 | 79.07 | 92.98 | 87.78 | 75.98 | 59.83 | 78.78 | 92.98 | 65.25 | 91.98 | 95.33 | 94.53 | 69.18 | 83.7 | 80.6 | 74.98 | 44.68 | 59.87 | 77.37 | 75.21 | 92.45 | 86.45 |
Model | Precision | Model URL | Aggregate Score | Grammar (Avg.) | Knowledge (Avg.) | Reasoning (Avg.) | Math (Avg.) | Classification (Avg.) | AGREE | ANLI | ARC-Challenge | ARC-Easy | Belebele | CTKFacts | Czech News | Facebook Comments | GSM8K | Klokanek | Mall Reviews | MMLU | SQAD | Subjectivity | TruthfulQA | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Llama-3.1-Nemotron-70B-Instruct-HF | bfloat16 | https://docs.anthropic.com/en/docs/about-claude/models#model-comparison-table | 68.45 | 75.92 | 86.05 | 62.13 | 47.56 | 78.27 | 75.92 | 41.58 | 88.05 | 96.97 | 91.62 | 49.46 | 84.9 | 81.9 | 74.45 | 20.67 | 65.23 | 74.93 | 65.84 | 89.55 | 83.13 |
Basic Information
The CzechBench evaluation suite is hosted on GitHub. It is implemented on top of the popular Language Model Evaluation Harness framework, which provides extensive model compatibility and optimal evaluation efficiency.
All currently supported benchmarks are listed in the table below:
| Dataset | Language | Task type | Metrics | Samples | Task ID |
|---|---|---|---|---|---|
| AGREE | CS (Original) | Subject-verb agreement | Acc | 627 | agree_cs |
| ANLI | CS (Translated) | Natural Language Inference | Acc, Macro F1 | 1200 | anli_cs |
| ARC Challenge | CS (Translated) | Knowledge-Based QA | Acc | 1172 | arc_cs |
| ARC Easy | CS (Translated) | Knowledge-Based QA | Acc | 2376 | arc_cs |
| Belebele | CS (Professional translation) | Reading Comprehension / QA | Acc | 895 | belebele_cs |
| CTKFacts | CS (Original) | Natural Language Inference | Acc, Macro F1 | 558 | ctkfacts_cs |
| Czech News | CS (Original) | News Topic Classification | Acc, Macro F1 | 1000 | czechnews_cs |
| Facebook Comments | CS (Original) | Sentiment Analysis | Acc, Macro F1 | 1000 | fb_comments_cs |
| GSM8K | CS (Translated) | Mathematical inference | EM Acc | 1319 | gsm8k_cs |
| Klokánek | CS (Original) | Math/Logical Inference | Acc | 808 | klokanek_cs |
| Mall Reviews | CS (Original) | Sentiment Analysis | Acc, Macro F1 | 3000 | mall_reviews_cs |
| MMLU | CS (Translated) | Knowledge-Based QA | Acc | 12408 | mmlu_cs |
| SQAD | CS (Original) | Reading Comprehension / QA | EM Acc, BoW F1 | 843 | sqad_cs |
| Subjectivity | CS (Original) | Subjectivity Analysis | Acc, Macro F1 | 2000 | subjectivity_cs |
| TruthfulQA | CS (Translated) | Knowledge-Based QA | Acc | 813 | truthfulqa_cs |
The leaderboard table also displays aggregated scores across task categories, including:
- Grammar (Avg.): AGREE
- Knowledge (Avg.): ARC-Challenge, ARC-Easy, MMLU, TruthfulQA
- Reasoning (Avg.): ANLI, Belebele, CTKFacts, SQAD
- Math (Avg.): GSM8K, Klokanek
- Classification (Avg.): Czech News, Facebook Comments, Mall Reviews, Subjectivity
- Aggregate Score: Average over above categories
Evaluation Process
1. Install CzechBench:
git clone https://github.com/jirkoada/czechbench_eval_harness.git
cd czechbench_eval_harness
pip install -e “.[api]”
2. Run evaluation
export MODEL=your_model_namewhere your_model_name is HF path for public model. For example:export MODEL=meta-llama/Meta-Llama-3.1-8B-Instructexport OUTPUT_PATH=my_output_pathwhere my_output_path is directory for evaluation reports
Run following command (you can adjust parameters like batch_size or device):
lm_eval --model hf \
--model_args pretrained=$MODEL \
--tasks czechbench_tasks \
--device cuda:0 \
--batch_size 1 \
--write_out \
--log_samples \
--output_path $OUTPUT_PATH \
--apply_chat_template \
For advanced usage instructions, please inspect the CzechBench README on GitHub or the official LM Evaluation Harness documentation.
3. Upload results to Leaderboard
Inside the $OUTPUT_PATH directory, you can find the file results.json.
To submit your evaluation results to our leaderboard, please visit the "Submit here!" section above and upload your results.json file.
✉️✨ Submit your model here!
What weight precision were you using during the evaluation?
