toorgil/mongolian-llm-benchmark

Mongolian LLM Benchmark

A combined Mongolian-language benchmark dataset for evaluating large language models. Aggregated from 19 community datasets on HuggingFace, normalized to a single unified schema.

Stat	Value
Total rows	47,974
Language	Mongolian (mn)
Source datasets	19
Question types	QA, MCQ, Problem Solving, Code, DPO, Instruction Following

---

Source datasets

Dataset	Rows	Category
TRUMO12/LLM_QA	10,000	general
joppari/mn_business_benchmark_dataset_medium	10,000	business
Asakuu/mongolian-mcq-dataset	2,719	multi-subject MCQ
baaska7/qa	2,154	history
Namka/mongoliin_nuuts_tovchoo	2,000	history
tugsukage/Physic_dataset	2,000	physics
Detsu2470/Geograpphic	2,000	geography
sanchirshde/omgggg	2,000	geography/science
boldoo1130/mongolian_sports_qa	2,000	sports
Sentry123/mongolian-astronomy-qa	2,000	astronomy
Mungu0/mongolian-animal-qa	1,717	animals
B211910811/physics-qa	1,840	physics
Raquellek/LLM_mn_dulguun	1,400	math / general
qqdqewekfk/mongolian-dataset	1,244	legal / civic
WrathX/uni-math1	1,000	math
iddno2306/dpo_redbook1000	1,000	animals (DPO)
usxkhuulll/mongolian-json-dataset	1,000	general
Kerosinn/mn_benchmark_dataset_dundangi	1,000	math / geometry
Bokhbat/Mongolian-LLM-Benchmark	900	culture, medical, code…

---

Schema

Each row has the following fields:

Field	Type	Description
id	string	Unique row identifier
source	string	Original HuggingFace repo
category	string	Domain / subject (e.g. physics, math, sports)
type	string	qa · mcq · problem_solving · dpo · code · instruction_following
instruction	string	Question or instruction text
input	string	Additional context / given data (empty string if none)
choices	list[str]	Answer choices for MCQ rows (empty list for others)
answer	string	Correct answer / expected output
thinking	string	Chain-of-thought solution steps (empty if unavailable)
difficulty	string	easy / medium / hard (empty if unavailable)
language	string	Always mn

---

Load examples

Load everything

from datasets import load_dataset

ds = load_dataset("toorgil/mongolian-llm-benchmark", split="train")
print(ds)
# Dataset({ features: [...], num_rows: 47974 })

Filter by category

ds = load_dataset("toorgil/mongolian-llm-benchmark", split="train")

physics = ds.filter(lambda x: x["category"] == "physics")
math    = ds.filter(lambda x: x["category"] == "math")
sports  = ds.filter(lambda x: x["category"] == "sports")

Filter MCQ rows only

mcq = ds.filter(lambda x: x["type"] == "mcq")

Pandas

import pandas as pd

df = ds.to_pandas()
print(df.groupby("category")["id"].count().sort_values(ascending=False))

Iterate with a map

def format_prompt(row):
    prompt = row["instruction"]
    if row["input"]:
        prompt += f"\n\n{row['input']}"
    if row["choices"]:
        for i, c in enumerate(row["choices"]):
            prompt += f"\n{'ABCD'[i]}) {c}"
    return {"prompt": prompt}

ds_with_prompts = ds.map(format_prompt)

---

Category breakdown (top 20)

Category	Rows
general	11,400
business	10,000
history	4,154
physics	4,139
animals	2,717
math	2,150
geography	2,100
sports	2,000
astronomy	2,000
геометр	500
culture	253
хими (chemistry)	250
элемент (elements)	250
механик (mechanics)	220
geometry	200
mathematics	200
chemistry	199
general_knowledge	150
medical	150
instruction_following	150

---

Benchmarking LLMs

The benchmarking.py script evaluates any LLM against this dataset and reports accuracy per category.

Install dependencies

pip install datasets anthropic openai   # install only the provider you use

Set your API key

export ANTHROPIC_API_KEY=sk-ant-...   # Anthropic
# or
export OPENAI_API_KEY=sk-...          # OpenAI

Run

# Benchmark claude-opus-4-5 on 50 samples per category
python benchmarking.py --model claude-opus-4-5 --samples 50

# Benchmark GPT-4o on 100 samples per category
python benchmarking.py --model gpt-4o --samples 100

# Benchmark only physics and math, save results
python benchmarking.py --model claude-haiku-4-5 \
    --categories physics,math,astronomy \
    --samples 200 \
    --output haiku_results.json

# Use a different judge model for free-form answers
python benchmarking.py --model gpt-4o-mini --judge-model gpt-4o --samples 50

Output example

══════════════════════════════════════════════════════════════════
  Model: claude-opus-4-5
══════════════════════════════════════════════════════════════════
Category                                  Score   Correct   Total  Errors
──────────────────────────────────────────────────────────────────
  astronomy                               88.0%        44      50       0
  physics                                 82.0%        41      50       0
  sports                                  78.0%        39      50       0
  math                                    74.0%        37      50       0
  geography                               72.0%        36      50       0
  history                                 68.0%        34      50       0
  general                                 65.0%        33      50       0
  animals                                 70.0%        35      50       0
  business                                71.0%        36      50       0
  ...
──────────────────────────────────────────────────────────────────
  OVERALL                                 74.2%       740    1000       0
══════════════════════════════════════════════════════════════════

Results saved → benchmark_results.json

Evaluation strategy

Row type	Evaluation method
mcq	Exact match on answer text or choice letter (A/B/C/D)
qa	LLM-as-judge — Mongolian judge prompt asks if meaning matches
problem_solving	LLM-as-judge
dpo	LLM-as-judge on chosen answer
code	LLM-as-judge
instruction_following	LLM-as-judge

benchmark_results.json structure

{
  "model": "claude-opus-4-5",
  "judge_model": "claude-opus-4-5",
  "dataset": "toorgil/mongolian-llm-benchmark",
  "samples_per_category": 50,
  "seed": 42,
  "overall": 0.742,
  "results": {
    "physics": { "correct": 41, "total": 50, "score": 0.82, "errors": 0 },
    "math":    { "correct": 37, "total": 50, "score": 0.74, "errors": 0 }
  }
}

---

Citation

If you use this dataset, please also cite the original source datasets listed above.

@dataset{mongolian_llm_benchmark_2025,
  title   = {Mongolian LLM Benchmark},
  author  = {toorgil},
  year    = {2025},
  url     = {https://huggingface.co/datasets/toorgil/mongolian-llm-benchmark},
  note    = {Aggregated from 19 community Mongolian datasets}
}