PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-SFT: Permissively-Licensed Multilingual Instruction-Following

Daniela Baiamonte, Kamyar Zeinalipour, Elena Fano, Stefano Simonazzi, Andrea Zugarini

arXiv:2610.22086v1cs.CLcs.AI

TL;DR

Existing English-centric SFT data can rely on proprietary-model outputs that restrict commercial use, while multilingual European resources remain comparatively limited. Multi-SFT assembles a permissively licensed, five-language collection using translation and regeneration, and reports stronger non-English performance than English-source data while largely preserving English performance and outperforming official instruct models in most multilingual cases.

  • Problem

    Existing English-centric SFT datasets may contain proprietary-model-generated content that restricts commercial use, while multilingual European resources remain comparatively limited.

    Source: §3 · p. 2, §3.1 · p. 3

  • Method

    Multi-SFT combines 502,818 permissively sourced instruction-response pairs across five European languages using translation and regeneration pipelines, with human evaluation of both.

    Source: §1 · p. 1

  • Results

    Across nine models and 13 benchmarks, Multi-SFT improves non-English performance over original English sources, largely preserves English performance, and beats official instruct models in most multilingual cases.

    Source: §1 · p. 1, §5 · pp. 6–7

  • Takeaways & Limitations

    Multi-SFT provides a permissively licensed foundation for building commercial multilingual LLMs.

    Source: §5 · p. 7

    The evaluation covers five European languages, leaving transfer to non-Latin-script and low-resource languages unclear.

    Source: Limitations · p. 7

Abstract

from arXiv · show

Supervised fine-tuning datasets remain overwhelmingly English-centric and commercially restricted, as their content is frequently generated by proprietary models whose terms of service forbid training competing systems. We introduce Multi-SFT, a commercially permissive collection of 502,818 instruction-response pairs across five European languages, built via translation and regeneration pipelines for tasks where translation is insufficient. Human evaluation yields inter-annotator agreement of at least 0.87 across all datasets and languages. Fine-tuning nine models from three different families (EuroLLM, OLMo 2, Qwen3; 1B-32B) and evaluating them on 13 benchmarks shows that Multi-SFT alone is competitive with larger augmented mixes and official instruction-tuned releases, with decisive advantages on safety tasks.

1 Introduction

Multi-SFT is a permissively licensed collection of 502,818 instruction-response pairs across five European languages, built with translation and regeneration pipelines. Across nine models and 13 benchmarks, it improves on English-source data in non-English languages while largely preserving English performance and outperforming official instruction-tuned models in most multilingual cases.

1 Introduction

  • 502,818 instruction-response pairs across five European languages form a collection of 13 heterogeneous datasets built exclusively from permissively licensed data and models.Table 1 compares SFT mixtures on dataset-license and data-provenance compliance.Source: §1 · p. 1, Table 1 · p. 1
  • Translation adapts existing English-centric datasets, while regeneration creates multilingual content for tasks where translation is insufficient, including conversation, constrained instruction following, and safety.Both pipelines receive systematic human evaluation, with inter-annotator agreement of at least 0.87 across datasets and languages.Source: §1 · p. 1
  • Multi-SFT outperforms official instruct models in most multilingual settings while remaining competitive on English and improving non-English performance over the original English sources.The collection largely preserves performance on English benchmarks.Source: §1 · p. 1
  • Fine-tuning nine models from EuroLLM, OLMo 2, and Qwen3 across 13 benchmarks validates the collection.Table 2 provides an overview of the multi-dataset collection and its sources.Source: §1 · p. 1, Table 2 · p. 2

2 Related Work

Multilingual European SFT resources remain comparatively limited, and proprietary-model provenance can restrict commercial use. Multi-SFT builds on translation and open-model generation while expanding task coverage, adding expert validation, and selectively regenerating non-compliant examples.

2 Related Work

  • High-quality multilingual SFT resources for European languages remain comparatively limited, while existing datasets often face restricted licensing or proprietary-model provenance concerns.Aya offers broad multilingual coverage but limited per-language depth and variable quality in automatically generated portions.Source: §2 · p. 2
  • EuroLLM, Tülu 3, Salamandra, and Apertus advance multilingual or open post-training, yet include components generated or processed with proprietary models.The passage names these works with their original citations: Martins et al. (2024), Lambert et al. (2024), Gonzalez-Agirre et al. (2025), and Apertus et al. (2025).Source: §2 · p. 2
  • Like the multilingual Smoltalk extension, this work uses translation and open-model generation, while expanding task coverage, adding expert-driven quality validation, and selectively regenerating non-compliant examples.The cited Smoltalk extension is Allal et al. (2025).Source: §2 · p. 2

3 Multi-SFT

Multi-SFT combines permissively licensed multilingual datasets with translation and synthetic-generation pipelines, applying human review to both. Its sampled evaluation reports generally high agreement and positive ratings, while identifying incomplete instructions and constraint violations as quality challenges.

3 Multi-SFT

  • Multi-SFT comprises 13 datasets and 502,818 examples across English, Italian, French, Spanish, and German, emphasizing diverse competences and tasks over generic volume.The source selection targets a heterogeneous mixture of tasks.Source: §3 · p. 2, §3.1 · p. 2

3.1 Dataset selection criteria

  • Inclusion requires commercial-use-compatible licensing and compliant synthetic-data provenance; 45 of 59 candidate datasets were excluded for licensing or provenance violations.Restricted-model examples in otherwise permissively licensed datasets were selectively regenerated with open-weight models.Source: §3.1 · p. 3

3.2 Translation Pipeline

  • Translation uses language-expert-designed prompts, blind review by a separate expert, and iterative revision until output quality is satisfactory.For multi-smol_summarize, source documents are translated before summaries are generated from scratch, with additional output-format variants.Source: §3.2 · p. 3

3.3 Synthetic Data Generation

  • Five synthetic datasets use expert-reviewed prompt design and iterative refinement, including safe-response red-teaming examples and culturally adapted multilingual dialogues.Examples include Multi-AdvBench and Multi-AyaRedTeaming, with generated pairs manually reviewed.Source: §3.3 · p. 3
  • Synthetic generation also covers multi-turn dialogues and constrained instruction following; Multi-PersonaIF modifies constraint sampling to prevent logically incompatible combinations.Multi-Dialogues adapts the SODAverse pipeline, while Multi-Safety uses internally authored harmful prompts and manually reviewed safe responses.Source: §3.3 · pp. 3–4

3.4 Annotation Process

  • For each dataset and language, 50 entries are sampled for annotation, independently assessed by two annotators; Multi-FLAN receives 60 per language across four sub-datasets.English is annotated only for datasets generated from scratch in English: Multi-PersonaIF and multi-smol_summarize.Source: §3.4 · p. 4, Table 3 · p. 4

3.5 Quality Assessment

  • Agreement uses quadratic-weighted Gwet’s AC2 on collapsed A/B, C, and D labels, chosen because skewed positive ratings can depress Cohen’s κ and the labels are ordinal.The quadratic weighting penalizes disagreements between non-adjacent categories more heavily.Source: §3.5 · p. 4
  • Agreement exceeds 0.90 in most dataset-language cases; Multi-FLAN French and German score 0.93 and 0.90, Multi-PersonaIF French 0.88, and Multi-TableGPT Spanish 0.87.Most entries receive positive ratings; Multi-PersonaIF has the highest badly-rated share, reaching 28% in Italian.Source: §3.5 · p. 4, Table 4 · p. 5

4 Experiments

Across nine models, Multi-SFT often beats official instruction-tuned models on multilingual averages while keeping pace with an expanded mixture and retaining much English performance. Results are compared across multilingual and English benchmarks.

4.1 Experimental Setup

  • Across the three model families, the nine base models span 1B–32B parameters and were trained on ML, the expanded FULL mixture, or English-only BL.They were evaluated on multilingual and English benchmarks, including 13 benchmark settings in total.Source: §4.1 · p. 5

4.2 Results

  • On multilingual averages, ML exceeds SFT† for 6 of 9 models, with gains across every EuroLLM and Qwen3 size.The largest reported gap is +15.9 on Qwen3 14B.Source: §4.2 · p. 5, Table 5 · p. 6
  • Using roughly one third of FULL’s data, ML keeps up with the larger mixture and beats it on 4 of 9 models.ML is within 1 point of FULL on Qwen3 8B and 14B.Source: §4.2 · p. 5, Table 5 · p. 6
  • On English benchmarks, ML matches or beats SFT† for 4 of 9 models; Qwen3 gains are +10.7 on 8B and +13.6 on 14B.OLMo 2’s SFT† recipe is better than ML, likely due to English-heavy pretraining.Source: §4.2 · p. 5
  • ML beats the English-only BL on multilingual benchmarks for 7 of 9 models, while BL leads on English for 7 of 9 with mostly modest gaps.BL uses the same source datasets as ML but keeps them in English.Source: §4.2 · p. 6
  • EuroLLM’s multilingual-average gains over SFT† are +0.6, +1.9, and +2.2 across its three scales, and it matches or exceeds SFT† on English.The reported Qwen3 multilingual-average gains are +8.6, +12.3, and +15.9; OLMo 2 trails SFT† by −0.1, −1.4, and −10.2.Source: §4.2 · p. 6

5 Conclusions

Multi-SFT is a commercially permissive instruction-following collection that addresses English-centricity and licensing constraints in existing SFT resources. Evaluations across nine models and 13 benchmarks show stronger non-English performance than the original English sources, while largely preserving English performance.

5 Conclusions

  • Multi-SFT contains 502,818 instruction-response pairs across five European languages and addresses English-centricity and provenance-constrained licensing.The languages are English, Italian, French, Spanish, and German.Source: §5 · p. 6
  • Across nine models and 13 benchmarks, Multi-SFT performs better on non-English languages than the original English sources containing non-permissive data from closed LLMs.It largely preserves English benchmark performance and outperforms official instruct models in most multilingual cases.Source: §5 · pp. 6–7
  • The collection provides a permissively licensed foundation for building commercial multilingual LLMs.Source: §5 · p. 7

Limitations

The paper notes limits in verifying third-party dataset provenance and in the scope and completeness of its human quality evaluation. Its language coverage also excludes settings whose linguistic properties differ substantially from the five selected European languages.

Limitations

  • Third-party dataset generation provenance cannot always be established with certainty, so compliance assessments rely on available documentation and are made to the best of the authors’ knowledge.The assessment draws on dataset cards and associated publications.Source: Limitations · p. 7
  • Human evaluation covers sampled subsets rather than every example, leaving quality issues in unsampled data not fully excluded.The finite annotator pool and structured schema may also miss quality dimensions relevant to downstream tasks.Source: Limitations · p. 7
  • The five-language scope leaves non-Latin-script, low-resource, and linguistically different settings unaddressed, and pipeline transfer to them is unclear.Permissive-source restrictions may also exclude high-quality datasets and limit coverage in some task domains.Source: Limitations · p. 7

A.1 Dataset Selection Statistics

The appendix reports how candidate datasets were screened, which models were selected for regeneration, and examples of regenerated instruction-response pairs across languages and tasks.

A.1 Dataset Selection Statistics

  • 45 of 59 surveyed candidate SFT datasets were excluded, chiefly for restrictive proprietary-model terms, incompatible licenses, or unverifiable provenance; 13 datasets remained after quality assurance.The survey excluded math and coding topics; 14 datasets initially remained before quality assurance.Source: §A.1 · p. 9

A.2 Regeneration Model Selection

  • Mistral-Small was selected for translation strength, especially outside English, while Qwen models generally performed better on text generation; all selected models allowed generated outputs to be reused for training.Selection also considered cross-language and task-format consistency and licensing conditions.Source: §A.2 · p. 9

A.3 Regenerated Examples

  • The regenerated-example appendix samples instruction-response pairs from 10 datasets across five languages, including summarization, table editing, dialogue, and recipe-writing tasks.Each example presents an instruction and its regenerated response.Source: §A.3 · pp. 9–11

A.4 Dataset Schema

The appendix describes dataset provenance metadata and the annotation procedures used to assess examples, including how agreement labels and annotator judgments were handled.

A.4 Dataset Schema

  • Each dataset entry can record its original source, regeneration model, and whether it was used as-is, translated, or regenerated.The passage directs readers to Table 6 for the complete schema.Source: §A.4 · p. 11

B Annotation Details

  • Annotation uses separate schemas for instruction-following examples and multi-dialogues, merges labels A and B for agreement calculations, and records instruction problems and error categories.Annotators were blind to whether examples came from translation or generation and to the model or prompt configuration.Source: B Annotation Details · p. 11

C Inter-Annotator Agreement: Detailed Analysis

  • The appendix examines datasets with lower inter-annotator agreement and investigates the sources of disagreement.The supplied passage states the analysis scope rather than its specific findings.Source: C Inter-Annotator Agreement: Detailed… · p. 11

C.1 Multi-FLAN: Synthetic Reasoning Ambiguity

The appendix examines sources of annotator disagreement in Multi-FLAN and Multi-PersonaIF, and language-specific errors in Multi-TableGPT. It highlights ambiguity in judging generated reasoning, instruction constraints, and translated multi-part prompts.

C.1 Multi-FLAN: Synthetic Reasoning Ambiguity

  • Multi-FLAN agreement was 0.93 in French and 0.90 in German; generated reasoning chains can be logically coherent yet stylistically artificial, creating genuine ambiguity in judging their quality.The paper frames this as task difficulty rather than annotation unreliability.Source: §C.1 · pp. 11–12

C.2 Multi-PersonaIF: Incomplete Instructions and Constraint Sensitivity

  • Multi-PersonaIF had agreement of 0.88 in French and C-rated entries reaching 25% in Italian and 18% in French; disagreements reflected both constraint enforcement and underspecified instructions.The paper identifies incomplete prompts as actionable through improved prompt design and post-hoc filtering or regeneration.Source: §C.2 · p. 12

C.3 Multi-TableGPT: Language-Specific Translation Challenges

  • Multi-TableGPT’s Spanish subset had a 24% C/D-rate, versus 2% in German, 2% in Italian, and 3% in French; errors often involved incomplete multi-step outputs or factual inaccuracies.The authors link these cases to instruction structure not being fully preserved during translation or generation, particularly for prompts requiring structured reasoning.Source: §C.3 · pp. 12–13

D Training Details

All runs use the same one-epoch training setup, with shared optimization settings and hardware-managed DeepSpeed ZeRO stages.

D Training Details

  • Every model trains for one epoch with seed 42 and mean loss reduction.Source: D Training Details · p. 13
  • Training uses a 2×10−5 learning rate with cosine decay, 3% linear warmup, no weight decay, and gradient clipping at 1.0.Source: D Training Details · p. 13
  • Runs use sequence length 4096, batch size 128, open-instruct, and DeepSpeed ZeRO stage 1, 2, or 3 as memory permits on eight NVIDIA H100 80 GB GPUs.Source: D Training Details · p. 13

Related papers

Loading 2610.22086v1…