Hugging Face's logo

Datasets:
alexantonov
/
chuvash_llm_testset

Modalities:
Text
Formats:
parquet
Languages:
Chuvash
Size:
< 1K
Libraries:
Datasets
pandas
License:
Dataset card Data Studio Files
xet
 Community
1
chuvash_llm_testset / README.md
alexantonov's picture
alexantonov
Update README.md
31ad025 verified
preview code
|
raw
history blame contribute delete
8.66 kB
metadata 
language:
  - cv
license: cc0-1.0
dataset_info:
  features:
    - name: number
      dtype: string
    - name: text
      dtype: string
    - name: question
      dtype: string
    - name: doc_number
      dtype: string
    - name: doc_name
      dtype: string
    - name: answer
      dtype: string
    - name: options
      sequence: string
    - name: answer_letter
      dtype: string
  splits:
    - name: train
      num_bytes: 53982
      num_examples: 100
  download_size: 31521
  dataset_size: 53982
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Header

Dataset Summary

Recent advances in LLMs have driven remarkable progress, yet their performance remains inconsistent on low-resource languages, highlighting challenges in equitable AI development.

This dataset demonstrates that while LLMs improve in Chuvash language understanding, fact-based knowledge about Chuvash literature remains a significant unresolved challenge.

The dataset is comprised of 100 questions designed to assess factual knowledge of Chuvash literature. The objective is to identify the name of a character based on a provided piece of text.

Models

  • Gemma: gemma-3-27b-it
  • Gemini: gemini-2.5-flash
  • Claude: claude-sonnet-4
  • GPT: gpt-4.1

Evaluation

To assess the quality, four approaches are utilized:

  • contains: an open-ended question.
  • contains_literature: an open-ended question which clarifies that the text is a piece from Chuvash literature.
  • contains_book: an open-ended question that also provides the title of the book the excerpt is from.
  • options: a question with 10 answer options.
def make_prompt_contains(text: str, question: str, lang: str) -> str:
  prompt = {
      "en": f"Text:\n\n{text}\n\nQuestion:\n\n{question}\n\n",
      "ru": f"Текст:\n\n{text}\n\nВопрос:\n\n{question}\n\n",
      "cv": f"Текст:\n\n{text}\n\nЫйту:\n\n{question}\n\n"
  }
  return prompt[lang]

def make_prompt_contains_literature(text: str, question: str, lang: str) -> str:
  prompt = {
      "en": f"Here is the text from british or american literature. Guess which book it's from and answer the question.\n\nText:\n\n{text}\n\nQuestion:\n\n{question}\n\n",
      "ru": f"Дан отрывок текста из русской литературы. Предположи, из какой книги и ответь на вопрос.\n\nТекст:\n\n{text}\n\nВопрос:\n\n{question}\n\n",
      "cv": f"Чӑваш литературин текст сыпӑкӗ панӑ. Текста хӑш кӗнекерен ҫырнине кала та ыйту ҫине ответле.\n\nТекст:\n\n{text}\n\nЫйту:\n\n{question}\n\n"
  }
  return prompt[lang]

def make_prompt_contains_book(text: str, question: str, book: str, lang: str) -> str:
  prompt = {
      "en": f"Here is the text from british or american literature. Based on book name please answer the question.\n\nBook name:{book}\n\nText:\n\n{text}\n\nQuestion:\n\n{question}\n\n",
      "ru": f"Дан отрывок текста из русской литературы. С учетом того, что название книги известно, ответь на вопрос.\n\nНазвание книги:{book}\n\nТекст:\n\n{text}\n\nВопрос:\n\n{question}\n\n",
      "cv": f"Чӑваш литературин текст сыпӑкӗ панӑ. Кӗнеке ячӗ паллӑ пулнине шута илсе ыйту ҫине ответле.\n\nКӗнеке ячӗ:{book}\n\nТекст:\n\n{text}\n\nЫйту:\n\n{question}\n\n"
  }
  return prompt[lang]

def convert_option_in_prompt_format(options: str) -> str:
    result = ''
    for i, option in enumerate(options):
        result += f'{chr(65 + i)}: {option}\n'
    return result.strip()

def make_prompt_options(text: str, question: str, options: str, lang: str) -> str:
  prompt_options = convert_option_in_prompt_format(options)

  prompt = {
      "en": f"Text:\n\n{text}\n\nQuestion:\n\n{question}\n\nSelect the correct answer from the options. It is IMPORTANT to write ONLY the answer letter in response:\n\n{prompt_options}\n\n",
      "ru": f"Текст:\n\n{text}\n\nВопрос:\n\n{question}\n\nВыбери правильный ответ из предложенных вариантов. ВАЖНО написать в ответе ТОЛЬКО букву ответа:\n\n{prompt_options}\n\n",
      "cv": f"Текст:\n\n{text}\n\nЫйту:\n\n{question}\n\nСӗннӗ вариантсенчен тӗрӗс хурав суйла. Хуравра хурав саспалли ҪЕҪ ҫырни ПӖЛТЕРӖШЛӖ:\n\n{prompt_options}\n\n"
  }
  return prompt[lang]

Languages

For the purpose of comparative analysis, we also gathered examples in high-resource languages: English and Russian (10 examples per language), all presented in an analogous format.

Additional dataset: chuvash_llm_testset_ru_en

Results

All model answers are opened: see result folder

File name looks like {model_name}.{eval_type}.{language}.txt ordered by number field

Best results

Best results were reached with contains_book evaluation type. See the figure at the top

Model Language True(%)
Gemma en 60
Gemini en 60
Claude en 90
GPT en 100
Gemma ru 60
Gemini ru 50
Claude ru 70
GPT ru 80
Gemma cv 2
Gemini cv 1
Claude cv 4
GPT cv 1

Other results

contains
Model Language True(%)
Gemma en 0
Gemini en 10
Claude en 40
GPT en 100
Gemma ru 0
Gemini ru 20
Claude ru 30
GPT ru 50
Gemma cv 1
Gemini cv 1
Claude cv 1
GPT cv 3
contains_literature
Model Language True(%)
Gemma en 40
Gemini en 70
Claude en 90
GPT en 100
Gemma ru 20
Gemini ru 60
Claude ru 70
GPT ru 70
Gemma cv 2
Gemini cv 4
Claude cv 4
GPT cv 1
options
Model Language True(%)
Gemma en 20
Gemini en 40
Claude en 60
GPT en 80
Gemma ru 0
Gemini ru 30
Claude ru 30
GPT ru 50
Gemma cv 0
Gemini cv 5
Claude cv 0
GPT cv 0

Code

Access to models via OpenRouter 

import requests
import json

OPENROUTER_API_KEY = "YOUR KEY"

def evaluate_openrouter(prompt: str, model_name: str) -> str:
  response = requests.post(
    url="https://openrouter.ai/api/v1/chat/completions",
    headers={
      "Authorization": f"Bearer {OPENROUTER_API_KEY}",
      ## "HTTP-Referer": f"{YOUR_SITE_URL}", # Optional, for including your app on openrouter.ai rankings.
      ## "X-Title": f"{YOUR_APP_NAME}", # Optional. Shows in rankings on openrouter.ai.
    },
    data=json.dumps({
      "model": model_name,
      "messages": [
        { "role": "user", "content": prompt }
      ]
    })
  )
  data = response.json()
  # print(data)
  return data["choices"][0]["message"]["content"]

Check correct 

def check_correct_contains(answer: str, llm_answer: str) -> bool:
  return answer in llm_answer

def check_correct_options(answer: str, llm_answer: str, options: str) -> bool:
  if llm_answer.isalpha() and 0 <= ord(llm_answer.lower()) - ord('a') < len(options):
    model_choice = options[ord(llm_answer.lower()) - ord('a')]
    return model_choice == answer
  else:
    return False

def check_correct(eval_mode: str, json_object: dict) -> bool:
    if eval_mode in ("contains", "contains_literature", "contains_book"):
        answer = json_object["answer"]
        llm_answer = json_object["llm_answer"]
        return check_correct_contains(answer, llm_answer)
    elif eval_mode == "options":
        answer = json_object["answer"]
        llm_answer = json_object["llm_answer"]
        options = json_object["options"]
        return check_correct_options(answer, llm_answer, options)
    else:
        raise ValueError("Invalid eval_mode: %s" % eval_mode)