henryhale/luganda-reasoning

Luganda Reasoning

Short question → reasoning → answer examples in Luganda, about half math and half commonsense, for teaching small models to reason step by step. Every row is at most 1,400 English characters before translation.

source	Taken from	Reasoning	Answer	Rows
openai/gsm8k	openai/gsm8k (main, train + test)	Worked solution	Number	TBD
ChilleD/StrategyQA	ChilleD/StrategyQA	Supporting facts	Yee / Nedda	2,035
tasksource/ecqa	tasksource/ecqa	Explanation of the choice	One of 5 options	TBD
cs-giung/clean-gsm8k-aug-nl	cs-giung/clean-gsm8k-aug-nl (random sample, no GSM8K repeats)	Step-by-step solution	Number	TBD

Usage

from datasets import load_dataset

ds = load_dataset("henryhale/luganda-reasoning", split="train")
train = ds.filter(lambda r: r["split"] != "test")  # keep the sources' test rows for evaluation
math = ds.filter(lambda r: "gsm8k" in r["source"])  # or pick sources by the `source` column

def to_text(r):
    return {"text": f"{r['question']}\n<think>\n{r['reasoning']}\n</think>\n{r['answer']}"}

train = train.map(to_text)

Columns: question, reasoning, answer, split (the split the row came from in its source: train / validation / test), source and source_id (<file>:<row> in the source repo).

How it was made

Translated from English with facebook/nllb-200-distilled-1.3B (CTranslate2, int8, beam 2), sentence by sentence. Worked arithmetic (48/2 = 24) and GSM8K answers were copied unchanged. A row was dropped if any translated sentence was empty, had a length ratio outside 0.5–2×, repeated itself or lost a number, or if a text of 50+ characters was not Luganda according to GlotLID. The translations were not reviewed by people. Expect some unnatural phrasing.

License

The translations were made with NLLB-200 (CC BY-NC 4.0), so use this dataset for non-commercial research only. Each row is also under its source's license: GSM8K, StrategyQA and clean-gsm8k-aug-nl are MIT, and ECQA is CDLA-Sharing-1.0. clean-gsm8k-aug-nl was written by GPT-4, so OpenAI's terms also apply to it.