PolyglotTeachers-SFT-Synth

This dataset contains synthetic supervised fine-tuning examples generated by the best teacher we found in the paper Polyglot Teachers: Evaluating Language Models for Multilingual Synthetic Data Generation, where we systematically characterize what makes a good teacher model. It contains examples across six languages: Arabic, Czech, German, Indonesian, Japanese, Spanish, and Tagalog. Note: In our experiments, we subsampled 10k examples per language for training. Here we release the full unfiltered set to enable reproducibility and give researchers the flexibility to construct their own subsamples or training mixtures.

Dataset Summary

• Languages: Arabic (ar), Czech (cs), German (de), Indonesian (id), Japanese (ja), Spanish (es), Tagalog (tl)
• Total examples: 315,596
• Teacher model: google/gemma-3-27b-it
• Generation strategies: generate, respond, translate

Language Distribution

Language	Examples
Indonesian (id)	85,952
German (de)	83,878
Arabic (ar)	77,770
Japanese (ja)	27,198
Tagalog (tl)	40,875
Spanish (es)	25,609
Czech (cs)	15,189

Data Sources

The seed data comes from several multilingual datasets, which were then used to synthesize new examples via Gemma-3-27B-IT. Each source dataset was processed using one of three strategies: generate (create new prompt-response pairs from a seed), respond (generate a response given a prompt), or translate (translate an English example into a target language).

• allenai/WildChat-4.8M: multilingual prompt-response pairs from real user interactions.
• openai/gsm8k: math word problems (English, translated into target languages).
• Magpie-Align/Magpie-Pro-300K-Filtered: general chat data (English, translated into target languages).
• nvidia/Helpsteer3: multilingual preference data.
• OpenAssistant/oasst2: multilingual assistant conversations.
• utter-project/EuroBlocks-SFT-Synthetic-1124: European multilingual synthetic data.
• CohereLabs/aya_collection: multilingual instruction data.
• arbml/CIDAR: Arabic instruction data.
• indonlp/cendol_collection_v2: Indonesian instruction data.

Dataset Structure

Each example contains the following fields:

Field	Type	Description
id	str	Unique identifier
source	str	Source dataset name
language	str	ISO 639-1 language code
strategy	str	Synthesis strategy used (generate, respond, or translate)
source_id	str	Identifier from the source dataset
synth_prompt	str	The prompt used to instruct the teacher model during synthesis
model	str	Teacher model used for generation
prompt	str	The user prompt
response	str	The model response
messages	list	Chat-formatted messages (role and content) for SFT

Usage

from datasets import load_dataset

ds = load_dataset("ljvmiranda921/PolyglotTeachers-SFT-Synth", split="train")

# Filter by language
arabic_ds = ds.filter(lambda x: x["language"] == "ar")

# Use the messages field directly for SFT
print(arabic_ds[0]["messages"])

Acknowledgements

LJVM and AK acknowledge the support of the UKRI Frontier Grant EP/Y031350/1 (EQUATE). This work was performed using joint resources provided by the Cambridge Service for Data Driven Discovery (CSD3) EP/T022159/1 and the Isambard AI National AI Research Resource (AIRR) ST/AIRR/I-A-I/1023, and the Microsoft Research Grant. LJVM would also like to thank Songbo Hu, Chen Cecilia Liu, Millicent Ochieng, and Felermino Ali for helpful and productive discussions on the project.

Citation

@misc{miranda2025polyglotteachers,
  title={{Polyglot Teachers: Evaluating Language Models for Multilingual Synthetic Data Generation}},
  author={Lester James V. Miranda and Ivan Vulić and Anna Korhonen},
  year={2025},
}