henryhale/luganda-sft-translated

Luganda SFT (machine-translated)

About 145k Luganda instruction/response pairs, machine-translated from five English instruction and dialogue datasets with facebook/nllb-200-distilled-600M (eng_Latn → lug_Latn). Nobody has reviewed the translations.

Fields

Field	Description
instruction_lug	Prompt, in Luganda
output_lug	Response, in Luganda
hf_dataset	Source dataset on the Hugging Face Hub (user/repo)

Sources

hf_dataset	Rows	What was translated	License
databricks/databricks-dolly-15k	15,011	instruction → response	CC BY-SA 3.0
HuggingFaceTB/smoltalk (everyday-conversations subset)	2,260	Last user turn → last assistant turn	Apache 2.0
yahma/alpaca-cleaned	51,760	instruction → output	CC BY-NC 4.0 (from the original Stanford Alpaca data)
pixelsandpointers/empathetic_dialogues_for_lm	~17.8k	Opening exchange only (conv[1] → conv[2])	CC BY-NC 4.0 (from the original EmpatheticDialogues)
pixelsandpointers/ed-for-lm	~58.8k	current → next (history not used)	CC BY-NC 4.0 (from the original EmpatheticDialogues)

License

Each row is under the license of its source, shown in the table above. Filter on source to keep only the licenses you can use. The translation model, NLLB-200, is itself CC BY-NC 4.0. Treat the dataset as non-commercial / research use only.

from datasets import load_dataset

ds = load_dataset("henryhale/luganda-sft", split="train")
ds = ds.filter(lambda r: r["hf_dataset"] == "HuggingFaceTB/smoltalk")

Limitations

• Translated by a small model and not reviewed, so expect errors and unnatural phrasing.
• Inputs were truncated at 512 tokens, so long texts may be cut short or missing sentences.
• Dolly's context field was not included. Rows from its closed QA, summarization and information-extraction categories refer to a passage that is missing.
• Code, math and markdown formatting may be garbled.
• empathetic and ed_turns come from the same conversations, so some pairs may appear in both.
• No deduplication or quality filtering was done.
• The content is translated English instruction data, not native Luganda text.

Credits

• Databricks Dolly 15k
• Hugging Face SmolTalk
• Stanford Alpaca (Taori et al., 2023), cleaned version from yahma/alpaca-cleaned
• EmpatheticDialogues, in versions from pixelsandpointers/empathetic_dialogues_for_lm and pixelsandpointers/ed-for-lm
• NLLB-200 by NLLB Team, translation model facebook/nllb-200-distilled-600M