Hugging Face's logo

Datasets:
tonychenxyz
/
mtob

Tasks:
Translation
Modalities:
Text
Formats:
json
Languages:
English
Size:
< 1K
ArXiv:
Tags:
translation
low-resource
kalamang
benchmark
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
prompt
stringlengths
754k
755k
category
stringclasses
2 values
extra_info
dict
"<|im_start|>system\nYou are a careful linguist that follows instructions and translates accurately (...TRUNCATED)
memwrap/mtob/ek
{"ground_truth":{"target":"Namana kat gerket?","source":"Who asked you?","direction":"ek","url":"htt(...TRUNCATED)
"<|im_start|>system\nYou are a careful linguist that follows instructions and translates accurately (...TRUNCATED)
memwrap/mtob/ek
{"ground_truth":{"target":"Karek mat kokarap.","source":"The rope is wrapped around it.","direction"(...TRUNCATED)
"<|im_start|>system\nYou are a careful linguist that follows instructions and translates accurately (...TRUNCATED)
memwrap/mtob/ek
{"ground_truth":{"target":"Wiseme mu toni Japangbon Beladarbon nausair.","source":"Before, they said(...TRUNCATED)
"<|im_start|>system\nYou are a careful linguist that follows instructions and translates accurately (...TRUNCATED)
memwrap/mtob/ek
{"ground_truth":{"target":"Davit emun pasa narawi.","source":"Davit's mother is filtering rice.","di(...TRUNCATED)
"<|im_start|>system\nYou are a careful linguist that follows instructions and translates accurately (...TRUNCATED)
memwrap/mtob/ek
{"ground_truth":{"target":"Mu me fam Yarkuran, Yorre ge.","source":"They are Yarkurans, not Yorres."(...TRUNCATED)
"<|im_start|>system\nYou are a careful linguist that follows instructions and translates accurately (...TRUNCATED)
memwrap/mtob/ek
{"ground_truth":{"target":"Goun yarnener ba goun rasa.","source":"He has a place on the stones but i(...TRUNCATED)
"<|im_start|>system\nYou are a careful linguist that follows instructions and translates accurately (...TRUNCATED)
memwrap/mtob/ek
{"ground_truth":{"target":"Kalis bara tong inggon se bo mikon.","source":"The rain falls, our barrel(...TRUNCATED)
"<|im_start|>system\nYou are a careful linguist that follows instructions and translates accurately (...TRUNCATED)
memwrap/mtob/ek
{"ground_truth":{"target":"An mat sirie ma bo met rep, ma koi kabasi rep.","source":"I ordered him t(...TRUNCATED)
"<|im_start|>system\nYou are a careful linguist that follows instructions and translates accurately (...TRUNCATED)
memwrap/mtob/ek
{"ground_truth":{"target":"Som yumene toni tumun yumene boda.","source":"That person says that kid i(...TRUNCATED)
"<|im_start|>system\nYou are a careful linguist that follows instructions and translates accurately (...TRUNCATED)
memwrap/mtob/ek
{"ground_truth":{"target":"Manado bo warta magaripta eba ma luk.","source":"Manado went fishing and (...TRUNCATED)
End of preview. Expand in Data Studio

MTOB for Code-LLaVA

Machine Translation of a low-resource language: Kalamang ↔ English translation benchmark.

Source

Dataset Structure

Two configurations are available:

  • memwrap: Grammar textbook wrapped in <|memory_start|> / <|memory_end|> markers
  • plain: Same content without memory markers (baseline)

Important: Only the grammar textbook is inside the memory block. Vocabulary list and instructions are outside.

Fields

  • prompt: Chat-templated prompt string
  • category: {variant}/mtob/{direction} where direction is ek (English→Kalamang) or ke (Kalamang→English)
  • extra_info:
    • ground_truth: Contains target, source, direction
    • scoring_function: mtob_translation

Usage 

from datasets import load_dataset

# Load memwrap variant
ds = load_dataset("tonychenxyz/mtob", "memwrap", split="test")

# Filter by direction
eng_to_kal = ds.filter(lambda x: "ek" in x["category"])
kal_to_eng = ds.filter(lambda x: "ke" in x["category"])

Evaluation

The scorer (mtob_translation) computes:

  • exact_match: Exact string match (case-insensitive, normalized)
  • chrf: Character n-gram F-score

Statistics

  • Translation pairs per direction (E→K and K→E)
  • Grammar textbook provides linguistic rules for Kalamang
  • Bilingual vocabulary for word-level translations
Downloads last month
15
Size of downloaded dataset files:
153 MB
Size of the auto-converted Parquet files:
82.8 MB
Number of rows:
200

Paper for tonychenxyz/mtob