id stringlengths 23 23 | task stringclasses 1
value | category stringclasses 1
value | text stringlengths 27 49 | answers listlengths 1 1 | has_error bool 1
class | explanation stringlengths 30 42 | script stringclasses 1
value | difficulty stringclasses 1
value | source stringlengths 29 44 | verified bool 1
class | license stringclasses 2
values |
|---|---|---|---|---|---|---|---|---|---|---|---|
srb-correction-c-c-0001 | srb_correction | c-ć | U razredu je dvadeset jedna vredna ućenica. | [
"U razredu je dvadeset jedna vredna učenica."
] | true | Treba „učenica“, a ne „ućenica“ (č/ć). | latin | hard | derived:srb-probe-broj-pridev-imenica-0009 | false | CC-BY-4.0 |
srb-correction-c-c-0002 | srb_correction | c-ć | Na ispitu je palo trideset ćetiri lenja studenta. | [
"Na ispitu je palo trideset četiri lenja studenta."
] | true | Treba „četiri“, a ne „ćetiri“ (č/ć). | latin | hard | derived:srb-probe-broj-pridev-imenica-0011 | false | CC-BY-4.0 |
srb-correction-c-c-0003 | srb_correction | c-ć | Ja uzeh kljuć sa stola i izađoh. | [
"Ja uzeh ključ sa stola i izađoh."
] | true | Treba „ključ“, a ne „kljuć“ (č/ć). | latin | hard | derived:srb-probe-aorist-imperfekat-syn-0006 | false | CC-BY-SA-4.0 |
srb-correction-c-c-0006 | srb_correction | c-ć | U ućionici su postavljena dva nova stola. | [
"U učionici su postavljena dva nova stola."
] | true | Treba „učionici“, a ne „ućionici“ (č/ć). | latin | hard | derived:srb-probe-broj-pridev-imenica-0001 | false | CC-BY-4.0 |
srb-correction-c-c-0007 | srb_correction | c-ć | Po staroj kuči noću su lutali duhovi. | [
"Po staroj kući noću su lutali duhovi."
] | true | Treba „kući“, a ne „kuči“ (č/ć). | latin | hard | derived:srb-probe-sibilarizacija-syn-0007 | false | CC-BY-SA-4.0 |
srb-correction-c-c-0008 | srb_correction | c-ć | Mi stigosmo na stanicu u poslednji ćas. | [
"Mi stigosmo na stanicu u poslednji čas."
] | true | Treba „čas“, a ne „ćas“ (č/ć). | latin | hard | derived:srb-probe-aorist-imperfekat-syn-0005 | false | CC-BY-SA-4.0 |
srb-correction-c-c-0009 | srb_correction | c-ć | Da li češ mi pomoći oko selidbe? | [
"Da li ćeš mi pomoći oko selidbe?"
] | true | Treba „ćeš“, a ne „češ“ (č/ć). | latin | hard | derived:srb-grammar-enklitike-0004 | false | CC-BY-4.0 |
srb-correction-c-c-0011 | srb_correction | c-ć | Ne boj se, mali vuće, ništa ti neće biti. | [
"Ne boj se, mali vuče, ništa ti neće biti."
] | true | Treba „vuče“, a ne „vuće“ (č/ć). | latin | hard | derived:srb-probe-sibilarizacija-syn-0020 | false | CC-BY-SA-4.0 |
srb-correction-c-c-0012 | srb_correction | c-ć | Dugo je prićala o svojoj nozi. | [
"Dugo je pričala o svojoj nozi."
] | true | Treba „pričala“, a ne „prićala“ (č/ć). | latin | hard | derived:srb-probe-sibilarizacija-syn-0014 | false | CC-BY-SA-4.0 |
srb-correction-c-c-0015 | srb_correction | c-ć | On mi tada reće da više ne dolazim. | [
"On mi tada reče da više ne dolazim."
] | true | Treba „reče“, a ne „reće“ (č/ć). | latin | hard | derived:srb-probe-aorist-imperfekat-syn-0009 | false | CC-BY-SA-4.0 |
srb-correction-c-c-0016 | srb_correction | c-ć | Bavi se fudbalom več deset godina. | [
"Bavi se fudbalom već deset godina."
] | true | Treba „već“, a ne „več“ (č/ć). | latin | hard | derived:srb-cloze-padezi-0004 | false | CC-BY-4.0 |
srb-correction-c-c-0017 | srb_correction | c-ć | U zoološkom vrtu ima ćetrnaest mladih lavova. | [
"U zoološkom vrtu ima četrnaest mladih lavova."
] | true | Treba „četrnaest“, a ne „ćetrnaest“ (č/ć). | latin | hard | derived:srb-probe-broj-pridev-imenica-0015 | false | CC-BY-4.0 |
srb-correction-c-c-0020 | srb_correction | c-ć | Kad je pala noč, deca zaspaše kao zaklana. | [
"Kad je pala noć, deca zaspaše kao zaklana."
] | true | Treba „noć“, a ne „noč“ (č/ć). | latin | hard | derived:srb-probe-aorist-imperfekat-syn-0008 | false | CC-BY-SA-4.0 |
srb-correction-c-c-0022 | srb_correction | c-ć | Proćitao je jedanaest debelih knjiga. | [
"Pročitao je jedanaest debelih knjiga."
] | true | Treba „Pročitao“, a ne „Proćitao“ (č/ć). | latin | hard | derived:srb-probe-broj-pridev-imenica-0007 | false | CC-BY-4.0 |
srb-correction-c-c-0023 | srb_correction | c-ć | Da li ćeš doči na sastanak sutra ujutru? | [
"Da li ćeš doći na sastanak sutra ujutru?"
] | true | Treba „doći“, a ne „doči“ (č/ć). | latin | hard | derived:srb-grammar-pravopis-0003 | false | CC-BY-4.0 |
srb-correction-c-c-0024 | srb_correction | c-ć | Hajde, junaće, pomozi mi da podignem ovo! | [
"Hajde, junače, pomozi mi da podignem ovo!"
] | true | Treba „junače“, a ne „junaće“ (č/ć). | latin | hard | derived:srb-probe-sibilarizacija-syn-0017 | false | CC-BY-SA-4.0 |
srb-correction-c-c-0025 | srb_correction | c-ć | Treba mi pomoč oko selidbe. | [
"Treba mi pomoć oko selidbe."
] | true | Treba „pomoć“, a ne „pomoč“ (č/ć). | latin | hard | derived:srb-grammar-standard-0007 | false | CC-BY-4.0 |
srb-correction-c-c-0026 | srb_correction | c-ć | Javiču ti se u vezi s tim sutra. | [
"Javiću ti se u vezi s tim sutra."
] | true | Treba „Javiću“, a ne „Javiču“ (č/ć). | latin | hard | derived:srb-grammar-standard-0002 | false | CC-BY-4.0 |
srb-correction-c-c-0027 | srb_correction | c-ć | Bože, saćuvaj nas od nevolje! | [
"Bože, sačuvaj nas od nevolje!"
] | true | Treba „sačuvaj“, a ne „saćuvaj“ (č/ć). | latin | hard | derived:srb-probe-sibilarizacija-syn-0015 | false | CC-BY-SA-4.0 |
srb-correction-c-c-0028 | srb_correction | c-ć | Ja stigoh kući tek pred ponoč. | [
"Ja stigoh kući tek pred ponoć."
] | true | Treba „ponoć“, a ne „ponoč“ (č/ć). | latin | hard | derived:srb-probe-aorist-imperfekat-syn-0002 | false | CC-BY-SA-4.0 |
srb-correction-c-c-0029 | srb_correction | c-ć | Hajde, momće, pomozi mi da podignem ovo! | [
"Hajde, momče, pomozi mi da podignem ovo!"
] | true | Treba „momče“, a ne „momće“ (č/ć). | latin | hard | derived:srb-probe-sibilarizacija-syn-0018 | false | CC-BY-SA-4.0 |
serbeval: a Serbian-language LLM benchmark (official set v1.2)
Author: Djordje Markovic (@djolex999)
Code, task YAMLs, full leaderboard and methodology: https://github.com/djolex999/serbeval
serbeval tests what translated benchmarks cannot carry over into Serbian: case government, verbal aspect, the aorist and imperfect, sibilarization, numeral agreement, the ijekavian jat reflex, pravopis (together/apart, hyphenated compounds, prefixoids), č/ć, and Cyrillic/Latin consistency. It runs on lm-evaluation-harness 0.4.12.
This dataset is the official set v1.2: 351 items in 17 tasks, plus the
harder v2 passage-correction config srb_v2_passages (76 passages; see below).
v1.1 and v1.2 (2026-10-17) fixed gold answers and dropped 3 items; v1.2 checked every changed
gold against the text of Pravopis 2010 (tačka and page in the explanation). Both spellings are
accepted only where Pravopis 2010 and Klajn disagree. The history is under errata in
MANIFEST.json.
Verification
Every item has a status in MANIFEST.json:
| Status | Items | Meaning |
|---|---|---|
human |
61 | checked by a native speaker (blind baseline 29/30, plus the disputed items) |
consensus |
214 | fewer than 2 of the 4 strongest models disagreed with the gold |
derived |
76 | correction item built mechanically from an official item |
Model consensus is not human verification. It is labelled separately on purpose.
Gold answers come from three places:
- hand-written items;
- the srLex lexicon, using only forms attested in the corpus;
- source checks against Pravopis 2010 and Klajn.
Items where the authorities disagree (e.g. foto-aparat vs fotoaparat) were dropped rather than pinned to one side.
Tasks
| Config | Items | What it tests |
|---|---|---|
srb_correction_c_c |
21 | correction: one č/ć error, no cue |
srb_correction_ispravno |
27 | correction: clean sentences (false-alarm check) |
srb_correction_razmak_crtica |
28 | correction: one spacing/hyphen error, no cue |
srb_cyrillic_agreement |
20 | MCQ: the same question in Cyrillic and Latin (pairs) |
srb_grammar |
19 | MCQ: case government, aspect, agreement, clitics, standard usage |
srb_grammar_cloze |
18 | cloze: produce the inflected form |
srb_idioms |
19 | MCQ: meaning of Serbian idioms |
srb_pravopis_poluslozenice |
19 | cloze: polusloženice |
srb_pravopis_prefiksoidi |
17 | cloze: foreign prefixoids |
srb_pravopis_sastavljeno_rastavljeno |
24 | cloze: together / apart |
srb_pravopis_slozeni_pridevi |
22 | cloze: compound adjectives |
srb_probe_aorist_imperfekat |
20 | cloze: aorist and imperfect |
srb_probe_broj_pridev_imenica |
20 | cloze: numeral + adjective + noun agreement |
srb_probe_jat |
19 | cloze: ijekavian jat reflex |
srb_probe_poluslozenice |
19 | cloze: hyphenated compounds |
srb_probe_sastavljeno_rastavljeno |
19 | cloze: written together or apart |
srb_probe_sibilarizacija |
20 | cloze: sibilarization / palatalization |
Three item shapes:
- MCQ:
question,choices(4),answer_index. Gold positions are balanced. - Cloze:
questionwith____,cue, andanswers, the list of every accepted form. Scoring folds case, outer punctuation and script, but never diacritics or an intra-word hyphen. - Correction:
text, which may contain one error;answers(the corrected sentence, or the text itself whenhas_erroris false). There is no cue.
All items also carry id, task, category, explanation, script, difficulty,
source, verified and license.
Usage
from datasets import load_dataset
ds = load_dataset("djolex999/serbeval", "srb_probe_aorist_imperfekat", split="test")
To score models with the exact prompts and scoring rules, use the task YAMLs from the GitHub repo:
lm_eval --tasks srb_official --include_path lm_eval_tasks ...
Leaderboard (official set v1.2, 30 models, zero-shot, runs 2026-10-12 to 17)
| # | Model | MCQ | Cloze | Pravopis | Probni | Ispravljanje | Prosek | max unparsed |
|---|---|---|---|---|---|---|---|---|
| 1 | openai_gpt-6-sol | 100% | 100% | 99% | 96% | 99% | 99% | 0.00 |
| 2 | google_gemini-3.8-flash | 100% | 100% | 95% | 97% | 98% | 98% | 0.00 |
| 3 | anthropic_claude-opus-5.5 | 100% | 100% | 95% | 96% | 99% | 98% | 0.00 |
| 4 | openai_gpt-6.1-sol | 100% | 100% | 94% | 97% | 99% | 98% | 0.00 |
| 5 | openai_gpt-5.6-sol | 100% | 100% | 92% | 97% | 98% | 97% | 0.00 |
| 6 | deepseek_deepseek-v4-pro | 98% | 100% | 89% | 90% | 95% | 95% | 0.00 |
| 7 | openai_gpt-6-luna | 100% | 100% | 83% | 96% | 94% | 94% | 0.00 |
| 8 | anthropic_claude-sonnet-5.5 | 96% | 94% | 92% | 92% | 96% | 94% | 0.00 |
| 9 | x-ai_grok-4.7 | 100% | 89% | 89% | 95% | 99% | 94% | 0.00 |
| 10 | google_gemini-3.5-flash-lite | 100% | 100% | 88% | 87% | 95% | 94% | 0.00 |
| 11 | openai_gpt-5.6-luna | 100% | 100% | 82% | 89% | 96% | 93% | 0.00 |
| 12 | moonshotai_kimi-k3 | 96% | 100% | 86% | 85% | 92% | 92% | 0.00 |
| 13 | deepseek_deepseek-v4.1-flash | 100% | 100% | 79% | 85% | 89% | 90% | 0.00 |
| 14 | z-ai_glm-5.3 | 98% | 94% | 82% | 85% | 93% | 90% | 0.05 |
| 15 | google_gemma-4-31b-it | 98% | 100% | 76% | 79% | 93% | 89% | 0.00 |
| 16 | deepseek_deepseek-v3.2 | 98% | 89% | 88% | 76% | 87% | 88% | 0.00 |
| 17 | minimax_minimax-m3 | 96% | 100% | 73% | 77% | 89% | 87% | 0.00 |
| 18 | qwen_qwen3.6-27b | 96% | 89% | 76% | 60% | 94% | 83% | 0.00 |
| 19 | mistralai_mistral-large-4-0 | 98% | 89% | 69% | 70% | 86% | 82% | 0.00 |
| 20 | qwen_qwen3-235b-a22b-2507 | 88% | 100% | 72% | 67% | 80% | 81% | 0.05 |
| 21 | openai_gpt-5-mini | 90% | 89% | 66% | 62% | 87% | 79% | 0.00 |
| 22 | anthropic_claude-haiku-5.5 | 93% | 89% | 59% | 62% | 89% | 78% | 0.00 |
| 23 | nvidia_nemotron-3-ultra-550b-a55b | 91% | 94% | 54% | 63% | 82% | 77% | 0.00 |
| 24 | meta-llama_llama-3.3-70b-instruct | 88% | 94% | 57% | 52% | 75% | 73% | 0.05 |
| 25 | google_gemma-3-12b-it | 79% | 89% | 49% | 44% | 75% | 67% | 0.00 |
| 26 | qwen_qwen3.5-9b | 86% | 83% | 47% | 37% | 76% | 66% | 0.00 |
| 27 | nvidia_nemotron-3.5-lightning | 81% | 39% | 12% | 14% | 54% | 40% | 0.00 |
| 28 | mistralai_mistral-nemo | 69% | 33% | 12% | 15% | 68% | 39% | 0.00 |
| 29 | meta-llama_llama-3.1-8b-instruct | 48% | 50% | 23% | 21% | 54% | 39% | 0.00 |
| 30 | cohere_command-a-plus | 74% | 33% | 19% | 28% | 37% | 38% | 0.15 |
The frontier is near ceiling (GPT-6 Sol 99%, three more at 98%); its misses are mostly usage-over-norm spellings. Below it, the set separates models from 38% to 95%, and pravopis separates them most. The best open-weights model is DeepSeek V4 Pro (95%); turning reasoning on does not raise any tested model by more than 2 points. Details and caveats, including a possible small selection bias toward the four consensus models, are in the GitHub repo.
v2: passage correction (srb_v2_passages, harder)
v1 saturated at the frontier, so v2 asks for whole passages: 76 human-verified passages
(5–7 sentences, Latin and Cyrillic), 218 errors in total, 18 passages with no error at
all. The model gets text with no hint of how many errors there are or what kind, and must
return the corrected passage. The errors sit where usage and the norm part ways: polusloženice
and compounds (ekspres lonac, kontakt osoba, spomen-ploča), together/apart (na brzinu,
bilo gde, usput), compound adjectives, hteo – ne hteo, č/ć. Every gold form cites
Pravopis 2010 (tačka/page) in pravopis_2010 (list of form, citation). 45 passages
were drafted for serbeval with an LLM and checked by a native speaker; 31 are official
texts of Serbian public bodies (not copyrighted, ZASP art. 6), many with the institution's own
real errors.
Fields: text (model input), gold (corrected passage), sites (each error: wrong,
accepted right forms, and free when both forms are acceptable — no credit, no penalty),
markup ({right|wrong} inline), origin, provenance, license.
Scoring (GitHub: serbeval.passage): headline is the share of passages corrected exactly
(case, punctuation and script folded; diacritics and intra-word hyphens never); per-edit recall
and precision are diagnostics, and precision catches over-correction.
v2 results (first run, 2026-10-22, zero-shot)
| Model | v1.2 (items correct) | v2: errors fixed | v2: passages fully correct |
|---|---|---|---|
| Claude Opus 5.5 | 98% | 89% | 74% |
| GPT-6 Sol | 99% | 86% | 72% |
| DeepSeek V4 Pro | 95% | 75% | 61% |
| DeepSeek V4.1 Flash | 90% | 72% | 55% |
| DeepSeek V3.2 | 88% | 75% | 50% |
| GLM-5.3 | 90% | 69% | 50% |
| Kimi K3 | 92% | 70% | 49% |
| Qwen3.6 27B | 83% | 58% | 41% |
| Qwen3 235B | 81% | 59% | 38% |
| Gemma 3 12B | 67% | 46% | 34% |
| Llama 3.3 70B | 73% | 56% | 30% |
| Nemotron 3 Ultra | 77% | 53% | 30% |
| Gemma 4 31B | 89% | 63% | 29% |
| MiniMax M3 | 87% | 40% | 28% |
| Qwen3.5 9B | 66% | 54% | 28% |
| Mistral Nemo | 39% | 38% | 21% |
| Llama 3.1 8B | 39% | 37% | 13% |
| Nemotron 3.5 Lightning | 40% | 25% | 13% |
| Model | Passage exact | Error passages exact | Clean left intact | Errors fixed (recall) | Precision | False alarms | Unparsed |
|---|---|---|---|---|---|---|---|
| anthropic/claude-opus-5.5 | 56/76 | 43/58 | 13/18 | 89% (193/218) | 96% | 7 | 0 |
| openai/gpt-6-sol | 55/76 | 38/58 | 17/18 | 86% (188/218) | 99% | 1 | 0 |
| deepseek/deepseek-v4-pro | 46/76 | 28/58 | 18/18 | 75% (163/218) | 99% | 1 | 0 |
| deepseek/deepseek-v4.1-flash | 42/76 | 24/58 | 18/18 | 72% (157/218) | 100% | 0 | 0 |
| z-ai/glm-5.3 | 38/76 | 22/58 | 16/18 | 69% (150/218) | 97% | 5 | 0 |
| deepseek/deepseek-v3.2 | 38/76 | 24/58 | 14/18 | 75% (164/218) | 90% | 19 | 0 |
| moonshotai/kimi-k3 | 37/76 | 25/58 | 12/18 | 70% (153/218) | 95% | 8 | 0 |
| qwen/qwen3.6-27b | 31/76 | 15/58 | 16/18 | 58% (126/218) | 95% | 7 | 0 |
| qwen/qwen3-235b-a22b-2507 | 29/76 | 17/58 | 12/18 | 59% (128/218) | 91% | 13 | 0 |
| google/gemma-3-12b-it | 26/76 | 10/58 | 16/18 | 46% (101/218) | 93% | 8 | 0 |
| nvidia/nemotron-3-ultra-550b-a55b | 23/76 | 13/58 | 10/18 | 53% (116/218) | 80% | 29 | 0 |
| meta-llama/llama-3.3-70b-instruct | 23/76 | 10/58 | 13/18 | 56% (121/218) | 75% | 40 | 0 |
| google/gemma-4-31b-it | 22/76 | 11/58 | 11/18 | 63% (137/218) | 91% | 13 | 0 |
| qwen/qwen3.5-9b | 21/76 | 9/58 | 12/18 | 54% (118/218) | 87% | 17 | 0 |
| minimax/minimax-m3 | 21/76 | 5/58 | 16/18 | 40% (88/218) | 88% | 12 | 1 |
| mistralai/mistral-nemo | 16/76 | 4/58 | 12/18 | 38% (82/218) | 66% | 42 | 1 |
| nvidia/nemotron-3.5-lightning | 10/76 | 2/58 | 8/18 | 25% (55/218) | 49% | 58 | 1 |
| meta-llama/llama-3.1-8b-instruct | 10/76 | 4/58 | 6/18 | 37% (80/218) | 52% | 75 | 13 |
| cohere/command-a-plus | 9/76 | 1/58 | 8/18 | 22% (48/218) | 37% | 82 | 21 |
The first table compares like with like: per error, frontier models fix ~86–89% on v2 against ~98–99% item accuracy on v1.2, i.e. roughly 5–10× more mistakes. "Passages fully correct" falls further (72–74%) because it is all-or-nothing over ~4 errors per passage. The errors sit where usage diverges from the norm (kontakt-osoba, ekspres-lonac, nabrzinu, do viđenja), so v2 measures norm adherence in those spots, not general Serbian ability. Command A+'s API returned empty answers, so its row is not reliable.
License
Items carry their own license field: 303 are CC BY 4.0
(hand-written for this project) and 48 are CC BY-SA 4.0
(word forms from srLex). In srb_v2_passages, handwritten passages are CC BY 4.0 and official
texts are not copyrighted works (Serbian Copyright Act, art. 6). The dataset as a whole is
distributed under CC BY-SA 4.0.
Attribution:
- srLex 1.3: Ljubešić, Nikola. CLARIN.SI. CC BY-SA 4.0. https://www.clarin.si/repository/xmlui/handle/11356/1233
- UD_Serbian-SET: Samardžić, Tanja; Miletić, Aleksandra; Ljubešić, Nikola. CC BY-SA 4.0.
Citation
@misc{serbeval2026,
title = {serbeval: a Serbian-language benchmark for large language models},
author = {Markovic, Djordje},
note = {GitHub: @djolex999},
year = {2026},
url = {https://github.com/djolex999/serbeval}
}
- Downloads last month
- 232