Datasets:
variant large_stringclasses 2
values | model_key large_stringclasses 7
values | model_name large_stringclasses 7
values | benchmark large_stringclasses 6
values | benchmark_name large_stringclasses 6
values | axis large_stringclasses 6
values | metric large_stringclasses 3
values | n int64 500 2.91k | score float64 2.4 78.1 | ci95_low float64 1.38 75.6 | ci95_high float64 4.15 80.3 | macro_f1 float64 13.6 78.1 ⌀ | mc2 float64 38.9 47.2 ⌀ | mc2_ci95_low float64 37.3 45.3 ⌀ | mc2_ci95_high float64 40.5 49.2 ⌀ | f1 float64 1.95 73.2 ⌀ | precision float64 50 73.1 ⌀ | recall float64 1 99.8 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
published | cosmos_gemma_9b | CosmosGemma-9B | mmlu | MMLU-TR | knowledge | accuracy | 2,905 | 50.981067 | 49.163108 | 52.796435 | 50.768179 | null | null | null | null | null | null |
published | trendyol_7b_chat | Trendyol-7B-Chat v0.1 | mmlu | MMLU-TR | knowledge | accuracy | 2,905 | 23.442341 | 21.937502 | 25.017324 | 13.618427 | null | null | null | null | null | null |
published | trendyol_7b_chat_v4 | Trendyol-7B-Chat v4.1 | mmlu | MMLU-TR | knowledge | accuracy | 2,905 | 40.309811 | 38.539999 | 42.105217 | 38.460185 | null | null | null | null | null | null |
published | wiroai_9b | WiroAI-9B | mmlu | MMLU-TR | knowledge | accuracy | 2,905 | 45.197935 | 43.395651 | 47.012901 | 43.324974 | null | null | null | null | null | null |
published | wiroai_8b | WiroAI-8B | mmlu | MMLU-TR | knowledge | accuracy | 2,905 | 42.822719 | 41.033977 | 44.630418 | 42.633683 | null | null | null | null | null | null |
published | kocdigital_8b | KoçDigital-8B v0.1 | mmlu | MMLU-TR | knowledge | accuracy | 2,905 | 37.693632 | 35.948691 | 39.471076 | 34.472287 | null | null | null | null | null | null |
published | kumru_2b_base | Kumru-2B-Base | mmlu | MMLU-TR | knowledge | accuracy | 2,905 | 24.096386 | 22.57606 | 25.685129 | 22.976364 | null | null | null | null | null | null |
published | cosmos_gemma_9b | CosmosGemma-9B | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 78.06341 | 75.599226 | 80.343446 | 78.107765 | null | null | null | null | null | null |
published | trendyol_7b_chat | Trendyol-7B-Chat v0.1 | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 23.479006 | 21.13657 | 25.99547 | 13.862933 | null | null | null | null | null | null |
published | trendyol_7b_chat_v4 | Trendyol-7B-Chat v4.1 | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 68.894602 | 66.180276 | 71.484943 | 68.839817 | null | null | null | null | null | null |
published | wiroai_9b | WiroAI-9B | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 73.86461 | 71.268392 | 76.304231 | 73.718257 | null | null | null | null | null | null |
published | wiroai_8b | WiroAI-8B | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 63.496144 | 60.69384 | 66.209888 | 63.498322 | null | null | null | null | null | null |
published | kocdigital_8b | KoçDigital-8B v0.1 | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 54.241645 | 51.374045 | 57.081412 | 53.460937 | null | null | null | null | null | null |
published | kumru_2b_base | Kumru-2B-Base | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 23.393316 | 21.054227 | 25.906995 | 22.040241 | null | null | null | null | null | null |
published | cosmos_gemma_9b | CosmosGemma-9B | hellaswag | HellaSwag-TR | commonsense | accuracy | 1,000 | 54.7 | 51.602617 | 57.761411 | 54.62979 | null | null | null | null | null | null |
published | trendyol_7b_chat | Trendyol-7B-Chat v0.1 | hellaswag | HellaSwag-TR | commonsense | accuracy | 1,000 | 34.8 | 31.91094 | 37.805393 | 34.748327 | null | null | null | null | null | null |
published | trendyol_7b_chat_v4 | Trendyol-7B-Chat v4.1 | hellaswag | HellaSwag-TR | commonsense | accuracy | 1,000 | 50 | 46.90696 | 53.09304 | 49.88323 | null | null | null | null | null | null |
published | wiroai_9b | WiroAI-9B | hellaswag | HellaSwag-TR | commonsense | accuracy | 1,000 | 51.9 | 48.801915 | 54.983544 | 51.886589 | null | null | null | null | null | null |
published | wiroai_8b | WiroAI-8B | hellaswag | HellaSwag-TR | commonsense | accuracy | 1,000 | 48.8 | 45.71244 | 51.896745 | 48.732369 | null | null | null | null | null | null |
published | kocdigital_8b | KoçDigital-8B v0.1 | hellaswag | HellaSwag-TR | commonsense | accuracy | 1,000 | 44.5 | 41.446705 | 47.595389 | 44.423942 | null | null | null | null | null | null |
published | kumru_2b_base | Kumru-2B-Base | hellaswag | HellaSwag-TR | commonsense | accuracy | 1,000 | 41.4 | 38.385788 | 44.480032 | 41.384679 | null | null | null | null | null | null |
published | cosmos_gemma_9b | CosmosGemma-9B | gsm8k | GSM8K-TR | math reasoning | exact_match | 500 | 72.8 | 68.736739 | 76.515591 | null | null | null | null | null | null | null |
published | trendyol_7b_chat | Trendyol-7B-Chat v0.1 | gsm8k | GSM8K-TR | math reasoning | exact_match | 500 | 3.2 | 1.979151 | 5.134487 | null | null | null | null | null | null | null |
published | trendyol_7b_chat_v4 | Trendyol-7B-Chat v4.1 | gsm8k | GSM8K-TR | math reasoning | exact_match | 500 | 61.6 | 57.263883 | 65.759233 | null | null | null | null | null | null | null |
published | wiroai_9b | WiroAI-9B | gsm8k | GSM8K-TR | math reasoning | exact_match | 500 | 40.6 | 36.383044 | 44.960293 | null | null | null | null | null | null | null |
published | wiroai_8b | WiroAI-8B | gsm8k | GSM8K-TR | math reasoning | exact_match | 500 | 40.6 | 36.383044 | 44.960293 | null | null | null | null | null | null | null |
published | kocdigital_8b | KoçDigital-8B v0.1 | gsm8k | GSM8K-TR | math reasoning | exact_match | 500 | 43.6 | 39.31856 | 47.979031 | null | null | null | null | null | null | null |
published | kumru_2b_base | Kumru-2B-Base | gsm8k | GSM8K-TR | math reasoning | exact_match | 500 | 2.4 | 1.378131 | 4.147706 | null | null | null | null | null | null | null |
published | cosmos_gemma_9b | CosmosGemma-9B | truthfulqa | TruthfulQA-TR | reliability | mc1 | 500 | 52.4 | 48.020822 | 56.742581 | null | 47.243709 | 45.269032 | 49.20461 | null | null | null |
published | trendyol_7b_chat | Trendyol-7B-Chat v0.1 | truthfulqa | TruthfulQA-TR | reliability | mc1 | 500 | 45.4 | 41.087573 | 49.782571 | null | 44.797089 | 42.603685 | 47.016958 | null | null | null |
published | trendyol_7b_chat_v4 | Trendyol-7B-Chat v4.1 | truthfulqa | TruthfulQA-TR | reliability | mc1 | 500 | 48.2 | 43.850659 | 52.576789 | null | 43.582116 | 41.974501 | 45.19786 | null | null | null |
published | wiroai_9b | WiroAI-9B | truthfulqa | TruthfulQA-TR | reliability | mc1 | 500 | 49.4 | 45.039013 | 53.770136 | null | 43.952041 | 42.033863 | 45.863605 | null | null | null |
published | wiroai_8b | WiroAI-8B | truthfulqa | TruthfulQA-TR | reliability | mc1 | 500 | 47.4 | 43.059811 | 51.779835 | null | 44.639554 | 42.738785 | 46.458707 | null | null | null |
published | kocdigital_8b | KoçDigital-8B v0.1 | truthfulqa | TruthfulQA-TR | reliability | mc1 | 500 | 47.4 | 43.059811 | 51.779835 | null | 44.761867 | 42.887308 | 46.586126 | null | null | null |
published | kumru_2b_base | Kumru-2B-Base | truthfulqa | TruthfulQA-TR | reliability | mc1 | 500 | 36.8 | 32.688455 | 41.112828 | null | 38.873878 | 37.250474 | 40.536653 | null | null | null |
published | cosmos_gemma_9b | CosmosGemma-9B | hallucination | HalluVerse-M3 | hallucination detection | accuracy | 800 | 67 | 63.667191 | 70.170327 | null | null | null | null | 73.170732 | 61.85567 | 89.552239 |
published | trendyol_7b_chat | Trendyol-7B-Chat v0.1 | hallucination | HalluVerse-M3 | hallucination detection | accuracy | 800 | 49.75 | 46.294767 | 53.207623 | null | null | null | null | 1.95122 | 50 | 0.995025 |
published | trendyol_7b_chat_v4 | Trendyol-7B-Chat v4.1 | hallucination | HalluVerse-M3 | hallucination detection | accuracy | 800 | 52.5 | 49.035885 | 55.940221 | null | null | null | null | 67.8511 | 51.410256 | 99.751244 |
published | wiroai_9b | WiroAI-9B | hallucination | HalluVerse-M3 | hallucination detection | accuracy | 800 | 58.625 | 55.178876 | 61.988688 | null | null | null | null | 70.629991 | 54.896552 | 99.004975 |
published | wiroai_8b | WiroAI-8B | hallucination | HalluVerse-M3 | hallucination detection | accuracy | 800 | 64.25 | 60.868079 | 67.495723 | null | null | null | null | 66.510539 | 62.831858 | 70.646766 |
published | kocdigital_8b | KoçDigital-8B v0.1 | hallucination | HalluVerse-M3 | hallucination detection | accuracy | 800 | 67.625 | 64.30507 | 70.776475 | null | null | null | null | 63.572433 | 73.139159 | 56.218905 |
published | kumru_2b_base | Kumru-2B-Base | hallucination | HalluVerse-M3 | hallucination detection | accuracy | 800 | 50.125 | 46.667943 | 53.580863 | null | null | null | null | 15.64482 | 52.112676 | 9.20398 |
arc_corrected | trendyol_7b_chat_v4 | Trendyol-7B-Chat v4.1 | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 70.094259 | 67.40499 | 72.651672 | null | null | null | null | null | null | null |
arc_corrected | wiroai_9b | WiroAI-9B | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 75.321337 | 72.767306 | 77.709212 | null | null | null | null | null | null | null |
arc_corrected | wiroai_8b | WiroAI-8B | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 64.695801 | 61.909685 | 67.385485 | null | null | null | null | null | null | null |
arc_corrected | kocdigital_8b | KoçDigital-8B v0.1 | arc | ARC-TR | scientific reasoning | accuracy | 1,167 | 55.441302 | 52.57643 | 58.27047 | null | null | null | null | null | null | null |
Turkish LLM Multi-Axis Evaluation — Per-Item Results
These are the evaluation results behind the following paper. The release has per-item outcomes for every model, so the scores, confidence intervals and significance tests can all be recomputed without running a model.
M. T. Yılmaz, M. A. Pala, Y. Çakır, "A Multi-Axis Comparative Evaluation of Turkish Large Language Models across Knowledge, Reasoning, and Reliability," 2026 Innovations in Intelligent Systems and Applications Conference (ASYU), IEEE, 2026. doi:10.1109/ASYU71790.2026.11772376
| Code | github.com/TarikYil/turkish-llm-eval |
| Archive | Zenodo 10.5281/zenodo.24332272 (v2.0.0; all versions: 10.5281/zenodo.18861749) |
| Paper | IEEE Xplore 10.1109/ASYU71790.2026.11772376 |
Seven open-weight Turkish LLMs were evaluated on six benchmarks:
- Shared protocol. Every model uses the same scoring rules and the same fixed inference settings.
- Scoring. Multiple-choice tasks are scored by log-likelihood. Generation tasks use greedy decoding with seed 42.
- Quantization. All models are loaded in 4-bit NF4.
- Statistics. Scores carry 95% Wilson confidence intervals. Pairwise differences are tested with paired McNemar tests.
No benchmark text is included. Items are identified by SHA-256 hashes. The underlying datasets belong to their authors.
Main results
Primary metric per benchmark, in %. All models were evaluated on the same items.
| Model | MMLU-TR (Acc) |
ARC-TR (Acc) |
HellaSwag-TR (Acc) |
GSM8K-TR (EM) |
TruthfulQA-TR (MC1) |
HalluVerse-M3 (F1) |
|---|---|---|---|---|---|---|
| CosmosGemma-9B | 51.0 | 78.1 | 54.7 | 72.8 | 52.4 | 73.2 |
| Trendyol-7B-Chat v0.1 | 23.4 | 23.5 | 34.8 | 3.2 | 45.4 | 2.0 |
| Trendyol-7B-Chat v4.1 | 40.3 | 68.9 | 50.0 | 61.6 | 48.2 | 67.9 |
| WiroAI-9B | 45.2 | 73.9 | 51.9 | 40.6 | 49.4 | 70.6 |
| WiroAI-8B | 42.8 | 63.5 | 48.8 | 40.6 | 47.4 | 66.5 |
| KoçDigital-8B v0.1 | 37.7 | 54.2 | 44.5 | 43.6 | 47.4 | 63.6 |
| Kumru-2B-Base | 24.1 | 23.4 | 41.4 | 2.4 | 36.8 | 15.6 |
| Significant pairs (McNemar, p<0.05) | 20/21 | 20/21 | 18/21 | 17/21 | 9/21 | 15/21 |
Kumru-2B-Base is a base model and was evaluated with a raw prompt. On GSM8K-TR it used a 3-shot context. Treat its scores as a base-model reference, not a direct comparison with the instruction-tuned models.
Configurations
| Config | Rows | Contents |
|---|---|---|
scores (default) |
model × benchmark | Primary metric with its 95% Wilson CI. Also macro-F1 (multiple-choice tasks); MC2 with a bootstrap CI (TruthfulQA); F1, precision and recall (HalluVerse). variant is published or arc_corrected (see Known issue). |
item_results |
48104 | One row per model × scored item: correct (0/1), pred/gold (option index, number or 0/1 label, where stored), mc2 (TruthfulQA), correct_arc_fixed, item_sha256. |
pairwise_mcnemar |
6 × 21 | All model pairs per benchmark: scores, discordant counts, McNemar p-value, significant_05. |
models |
7 | Hugging Face ID, size, base model, training strategy, prompting, and whether predictions were stored. |
subset_manifest |
6874 | Every item in the evaluated subsets. Includes evaluated, eval_index, source, subset, flag, text_sha256 and item_sha256. |
item_results.eval_index matches subset_manifest.eval_index within each benchmark. Every model has exactly the same rows, so paired tests can be run directly.
Benchmarks
| Axis | Benchmark | Source | Items | Scoring |
|---|---|---|---|---|
| Knowledge | MMLU-TR (12 subjects) | malhajar/mmlu-tr |
2905 | log-likelihood of option letter |
| Scientific reasoning | ARC-TR (Challenge) | malhajar/arc-tr |
1167 | log-likelihood of option letter |
| Commonsense | HellaSwag-TR | malhajar/hellaswag-tr |
1000 | length-normalized log-likelihood of the ending |
| Math reasoning | GSM8K-TR | malhajar/gsm8k_tr-v0.2 |
500 | greedy generation (≤300 tokens), exact match on the final number |
| Reliability | TruthfulQA-TR | malhajar/truthfull_qa-tr |
500 | MC1 / MC2 with length-normalized log-likelihood |
| Hallucination | HalluVerse-M3 (Turkish QA) | sabdalja/HalluVerse-M3 |
800 (402 / 398) | greedy generation (≤8 tokens), DOĞRU/YANLIŞ; accuracy and F1 |
Each subset was drawn by shuffling with random.seed(42) and keeping the first N items. The positive class for HalluVerse-M3 is "hallucinated".
Usage
from datasets import load_dataset
scores = load_dataset("tarikk/turkish-llm-eval-results", "scores", split="train").to_pandas()
items = load_dataset("tarikk/turkish-llm-eval-results", "item_results", split="train").to_pandas()
# Example: paired comparison of two models on ARC-TR
arc = items[items.benchmark == "arc"].pivot(index="eval_index", columns="model_key", values="correct")
a, b = arc["cosmos_gemma_9b"], arc["wiroai_9b"]
print("A only:", ((a == 1) & (b == 0)).sum(), " B only:", ((a == 0) & (b == 1)).sum())
Known issue: ARC-TR answer keys given as digits
In the ARC-TR subset, 22 answer keys are digits ("1"–"4"). These keys are 1-based, but the evaluation code read them as 0-based. As a result:
- 20 scored items were compared against the wrong option;
- 2 items were skipped, which is why 1167 of the 1169 items are scored.
The affected items carry flag = arc_digit_key in subset_manifest. The column correct_arc_fixed re-scores them from the stored predictions. Where no predictions were stored, it is null.
| Model | Published | Corrected |
|---|---|---|
| CosmosGemma-9B | 78.1 | 78.0 – 79.6 (no stored predictions) |
| Trendyol-7B-Chat v0.1 | 23.5 | 23.4 – 25.0 (no stored predictions) |
| Trendyol-7B-Chat v4.1 | 68.9 | 70.1 |
| WiroAI-9B | 73.9 | 75.3 |
| WiroAI-8B | 63.5 | 64.7 |
| KoçDigital-8B v0.1 | 54.2 | 55.4 |
| Kumru-2B-Base | 23.4 | 22.9 – 24.1 (no stored predictions) |
ARC-TR scores change by at most about 1.5 percentage points. The number of significant ARC-TR pairs stays at 20/21 at both ends of the ranges, and the paper's conclusions are unaffected. Published scores are kept as variant = published.
Limitations
- Translated benchmarks. Five of the six benchmarks are Turkish translations of English sources. Contamination from pre-training cannot be ruled out.
- Quantization. All results are relative performance under shared 4-bit NF4 inference, not full-precision capability.
- Not bit-exact across runs. 4-bit GPU inference is not bit-exact across runs or environments. A repeated WiroAI-9B run with identical code differed on 4 of the 500 GSM8K-TR items.
- Observational comparison. The models differ in base family, size and training strategy all at once, so differences are comparative and not causal.
Citation
@inproceedings{yilmaz2026multiaxis,
author = {Yılmaz, Muhammet Tarık and Pala, Muhammed Ali and Çakır, Yusuf},
title = {A Multi-Axis Comparative Evaluation of Turkish Large Language Models across Knowledge, Reasoning, and Reliability},
booktitle = {2026 Innovations in Intelligent Systems and Applications Conference (ASYU)},
publisher = {IEEE},
year = {2026},
doi = {10.1109/ASYU71790.2026.11772376}
}
License
The results and this card are released under the MIT License, the same as the code. The benchmark datasets and the evaluated models are subject to their own licenses.
- Downloads last month
- -