No Language Left Behind: Scaling Human-Centered Machine Translation

• The project targets an unevenly served translation landscape in which most advances benefit high-resource languages, while low-resource languages face limited data and technical support.Only 25.9% of internet users speak English, yet 63.7% of websites are in English, illustrating the access gap for many low-resource communities.Source: §1 · p. 4
• NLLB organizes its effort around understanding native speakers, automatically creating training data, building translation models, and evaluating every target language.These four cornerstones are presented as the project’s low-resource translation strategy.Source: Figure 1 · p. 5
• NLLB addresses translation safety through toxic-word resources for all 200 languages and human evaluations designed to detect harmful hallucinated translations.The project combines automatic and human evaluations and reports qualitative insight into translation impact.Source: §1 · pp. 5–7
• The project creates Flores-200 for 40,602 translation directions, mines parallel data with LASER3, and trains multilingual models using mined and human-translated data.LASER3 identifies aligned bitext for 148 languages, while Flores-200 provides evaluation coverage in 204 languages.Source: §1 · pp. 5, 7
• The authors open source benchmarks, datasets, tools, scripts, and models, including a 54.5B Sparsely Gated Mixture-of-Experts model and distilled dense models, and deploy techniques to support Wikipedia editors.The released resources include NLLB-Seed, NLLB-MD, Toxicity-200, language-identification tools, LASER3, and models covering 202 languages.Source: §1 · pp. 6–7
• NLLB’s final model supports 202 languages and performs 40% better than the previous state of the art on Flores-101.The effort uses multilingual Mixtures-of-Experts models trained with mined data and human-translated seed data.Source: §1 · pp. 5, 7

2. Human-Centered Low-Resource Language Translation

• The interviews were designed to capture speakers’ needs, concerns, uses, and expectations for machine translation while centering community-level interests and ethical and social considerations.The project adopted a Value Sensitive Design orientation and began with the people most affected by NLLB.Source: §2 · p. 7, §2.1 · p. 8

2.1 Exploratory Interview Study Research Design

• The study interviewed 44 native speakers across 36 low-resource languages, defining low-resource languages as having fewer than 1 million publicly available example-translation sentences.Participants represented diverse regions and backgrounds; interviews were conducted remotely, lasted about 1.5 hours on average, and were systematically coded.Source: §2.1 · p. 8
• The authors acknowledge that recruiting diasporic participants and oversampling tech workers may bias perspectives, although participants also offered critical, reflexive views of translation’s risks.The authors argue that transnational ties and participants’ technological experience could help them speak knowledgeably about their language communities and identify potential harms.Source: §2.1 · pp. 8–9

2.1.1 Why should we prioritize low-resource languages?

• Participants linked native-language support to cultural identity, preservation, digital visibility, and more equitable access to knowledge, education, and employment.They described pressures from dominant languages and argued that machine translation could support bidirectional cultural learning and encourage institutions to invest in historically marginalized languages.Source: §2.1.1 · pp. 9–10
• Interviewees emphasized that safe, high-quality translation matters more than language inclusion alone because unreliable or toxic outputs discourage use and can erode confidence.Some participants reported that correcting machine output took longer than translating manually, while others sought technology that was reasonably usable rather than perfect.Source: §2.1.1 · p. 10

2.2 No Language Left Behind: Guiding Principles

• NLLB establishes four guiding principles: prioritize underserved communities, open-source contributions, work interdisciplinarily, and remain reflexive about judgments, practices, and trade-offs.These principles combine interview insights with responsible-AI literature and shape the project’s ethical, methodological, and documentation practices.Source: §2.2 · pp. 11–12

• The language-selection process narrowed over 250 candidates to around 210, expanding coverage from Flores-101’s 101 languages toward 200+ supported languages.Candidates were identified using Wikipedia presence, native-speaker regional input, existing NLP datasets, approximate speaker counts, and other community-level variables, with particular attention to historically underrepresented African languages.Source: §3 · pp. 12, 16
• The final catalog contains 204 languages, documented in Table 1 with language codes, names, scripts, families, resource levels, web support, and variant specifications.The catalog uses BCP 47 tags combining ISO 639-3 language codes with ISO 15924 script subtags, and supports languages represented in multiple scripts.Source: Table 1 · p. 16, §3 · p. 17
• For each candidate, linguists assessed standardized written forms because reliable evaluation datasets were considered critical for accelerated experimentation.This process also documented regional variation and multiple scripts, including for Fulah, Kashmiri, and Central Kanuri, but standardization can disadvantage languages with diverse writing practices.Source: §3 · p. 16
• Language inclusion ultimately prioritized potential impact on low-resource communities, while excluding languages with extremely few speakers or without written standards or predominantly oral modalities.The authors note that many selected languages have Wikipedia presence and come from historically underrepresented regions, while future work should address other modalities.Source: §3 · p. 17
• A language was classified as low-resource when it had fewer than 1M publicly available, de-duplicated bitext samples with any other language in the 200-language set, yielding 150 low-resource languages.The criterion includes non-English pairings, recognizing that regional datasets can provide resources through languages such as French in Africa.Source: §3 · p. 17

4.1.1 Benchmark Creation for Low-Resource Languages

• FLORES-200 quality control combines translator-reviewer alignment, staged translation, independent review, arbitration, and a 90% inclusion threshold.The workflow includes initial pilot translations, feedback on linguistic conventions, full translation, final sampling, and optional rework when quality falls below 90%.Source: Figure 4 · p. 19, §4.1.1 · pp. 20–21

4.1 FLORES-200

• FLORES-200 provides a 204-language, many-to-many benchmark of 3001 professionally translated sentences, doubling Flores-101’s language coverage.The benchmark uses content from Wikinews, Wikijunior, and Wikivoyage and supports reliable evaluation across 200+ languages.Source: §4.1.1 · p. 19, §4.1 · p. 19

4.1.2 Benchmark Creation for Non-English Directions

• Non-English workflows extend FLORES-200 through language-specific Arabic adaptation and script transliteration, using customized quality assessment and thresholds up to 95%.Arabic variants may be translated directly or adapted from Modern Standard Arabic with termlists, while four languages are transliterated from Latin-script counterparts.Source: §4.1.2 · p. 21

4.1.3 Flores-200 at a glance

• FLORES-200 reduced languages requiring re-translation from 45 in Flores-101 to 10, but low-resource languages still caused standardization, reviewer, and turnaround challenges.Mistranslation, unnatural translation, borrowing, literalness, and spelling inconsistency remained common quality issues; difficult languages such as Sicilian and Buginese took 287 days.Source: §4.1.3 · pp. 22–23

4.2 NLLB Seed Dataset

• NLLB-Seed supplies around 6,000 professionally translated Wikipedia-domain sentences in 39 languages for language identification, bitext training, and domain fine-tuning.Unlike FLORES-200, NLLB-Seed is intended for training rather than evaluation and therefore does not undergo the benchmark’s human quality-assurance process.Source: §4.2 · pp. 23–24
• NLLB-Seed is English-centric and may underrepresent non-English cultural content while introducing translationese into its human-translated targets.Its source material comes from English Wikipedia, reflecting the priorities of English Wikipedia editors rather than diverse cultural coverage.Source: §4.2 · p. 24

4.3 NLLB Multi-Domain Dataset

• NLLB-MD contains 3,000 professionally translated sentences in each of four domains across six language directions to test out-of-domain generalization and in-domain fine-tuning.The six languages include five low-resource languages and Russian, while the domains are news, scripted formal speech, unscripted informal speech, and health.Source: §4.3 · pp. 24–25

4.4 Conclusion

• FLORES-200 enables evaluation, NLLB-Seed supports encoder and translation-model training, and NLLB-MD measures translation generalizability across domains.The three datasets provide complementary resources for evaluating, creating, and adapting translation systems for languages with marginal prior support.Source: §4 · p. 18, §4.4 · p. 25

5. Automatically Creating Translation Training Data for Hundreds of Languages

• The pipeline extends scarce low-resource bitext by collecting monolingual web data, applying language identification and filtering, then mining likely aligned sentence pairs with LASER3.It is designed to address limited aligned and monolingual data across hundreds of languages and the noise of web-scale corpora.Source: §5 · pp. 25–26, Figure 6 · p. 26, Figure 7 · p. 27

5.1.1 Related Work

• LID must handle domain mismatch, severe class imbalance, noisy web text, short or mixed-language inputs, and confusion among similar languages while remaining efficient at web scale.Existing systems vary in coverage, but found-data approaches can conflate domain with language; the authors therefore tune on the uniform-domain Flores-200 development set.Source: §5.1 · pp. 26–27, §5.1.1 · p. 28

5.1.2 Models

• The LID system uses a fasttext character n-gram classifier with softmax, selected because it generalizes better out of domain and is computationally simpler than multiple binary classifiers.Temperature upsampling addresses severe class imbalance, including English at 10.1% of training data versus Minangkabau at 0.06%.Source: §5.1.2 · pp. 28–29

5.1.3 Improving LID with Linguistic Analysis

• Collaboration with linguists exposed spurious correlations and informed inspection tools and filters based on character distributions, expected scripts, and English detection.The inspection interface revealed phonotactic similarity and mutual intelligibility between Standard Malay and Indonesian, among the most frequently confused pairs.Source: §5.1.3 · pp. 29–31

5.1.4 Results

• Our LID model handles all 200 Flores-200 languages and significantly outperforms LangId, LangDetect, and CLD3 while achieving a lower false-positive rate.On human-annotated noisy web data, it also achieves higher F1 at lower FPR than CLD3, although performance is lower than on Flores-200 because of domain mismatch.Source: §5.1.4 · pp. 31, 33, Table 4 · p. 32, Table 5 · p. 32

5.2.1 Description of our Monolingual Pipeline

• The monolingual pipeline starts from CommonCrawl and ParaCrawl, applies paragraph-level LID to select language-specific sentence splitters, then reruns sentence-level LID to discard mixed-language, low-confidence, or wrong-script sentences.This end-to-end curation and cleaning process aims to produce high-quality monolingual data for bitext mining, backtranslation, and language-model training.Source: §5.2 · p. 34, §5.2.1 · pp. 34–35
• The monolingual pipeline combines language/script identification, quality heuristics, deduplication, and selective language-model filtering, while relaxing thresholds because filtering ratios vary across languages.URLs and hashtags are truncated before identification; high-resource languages use stronger filtering when sufficient data remain, whereas low-resource languages require more permissive thresholds.Source: §5.2.1 · pp. 35–36

5.2.2 Monolingual Data at a Glance

• 43.7 billion monolingual sentences remained after filtering 107.9 billion web paragraphs and discarding 2.8 billion sentences for language/script mismatch or low language-identification scores.The retained data were 90.8% high-resource and 9.2% low-resource, including 21.5 billion English sentences; high-resource data were further restricted to the top 30% by language-model score.Source: §5.2.2 · pp. 36–37

5.3 Mining Bitexts for Low-Resource Languages

• Bitext mining learns a multilingual sentence-embedding space and uses similarity margins to identify parallel sentences across monolingual collections, primarily pairing low-resource languages with English.The approach follows global mining, which compares candidate sentence pairs across collections rather than restricting alignment to preselected document pairs.Source: §5.3 · p. 37

5.3.1 Related work

• Prior scalable mining systems evolved from structure- and text-based alignment toward multilingual embedding spaces, while sentence-representation methods addressed weaknesses of generic multilingual pretraining for bitext retrieval.The paper situates its approach among ParaCrawl, CCMatrix, Samanantar, SBERT, LaBSE, LASER, and teacher–student methods for extending representations to new languages.Source: §5.3.1 · pp. 38–39

5.3.2 Student-Teacher Mining Approach

• The LASER2 teacher and language-specific LASER3 students add low-resource languages without retraining the full model, preserve compatible embedding spaces, and avoid competition with high-resource languages for capacity.Students are specialized by language or related-language group, may use dedicated vocabularies, minimize cosine loss, and can incorporate masked-language-modeling loss.Source: §5.3.2 · pp. 39–40
• The encoder-selection proxy xsim evaluates margin-based multilingual retrieval on Flores-200 devtest, after which mined bitext is tested by comparing NMT trained with mined data against the same system trained on existing bitext only.The end-to-end experiments use a margin threshold of 1.06, restrict translation to foreign-to-English directions, and mine from 21.5 billion unique English sentences.Source: §5.3.2 · pp. 40–41

5.3.3 Language-Specific Encoder Training

• LASER3 reduced average xsim error from 61 to 0.91, including reductions from 93.3 to 0.89 for Burmese and from 92.5 to 0.79 for Irish.For European minority languages, pairing each language with a similar higher-resource language enabled substantial mining; three exceeded BLEU 20 and two exceeded BLEU 30.Source: §5.3.3 · pp. 41–42, Figure 13 · p. 42
• Mining effectiveness depended strongly on available monolingual data: Papiamento reached BLEU 40.9 with over 7M mined bitexts, whereas Kabuverdianu reached BLEU 4.9 with fewer than 300k monolingual sentences.Javanese improved from BLEU 11.1 to 31.2, and Sundanese from 1.5 to 28.5, when languages were trained together in one encoder.Source: §5.3.3 · pp. 43–44
• Language-specific teacher-student encoders and vocabularies reduced xsim error to 0.1% for Amharic and 0.89% for Tigrinya, outperforming LaBSE for Tigrinya.The best model used Amharic and Tigrinya with English and an 8k SentencePiece vocabulary supporting the Ge’ez script.Source: §5.3.3 · p. 45

5.3.4 Mining at a Glance

• Mining produced 761 million English-paired sentence pairs across 148 bitexts and 302 million sentence pairs across 1465 non-English bitext pairs.The non-English collection includes pairings within African, Indic, and Malayo-Polynesian families and African languages with French.Source: §5.3.4 · p. 45

5.3.5 Limitations of Large-Scale Mining

• Mining quality is constrained by scarce monolingual data, filtering and alignment losses, uneven web coverage, and possible inclusion of previously machine-translated or duplicated content.These limitations can affect languages disproportionately, while mined data may also include content such as the Christian Bible from other sources.Source: §5.3.5 · pp. 45–46

5.3.6 Ethical Considerations for Mining Research

• Although stakeholders viewed language access as potentially necessary for advancement, the paper emphasizes that ownership of low-resource language data belongs to its speakers.The section distinguishes access from ownership when deploying low-resource language datasets.Source: §5.3.6 · p. 46

5.4 Conclusion

• Over 1.1 billion new sentence pairs were created for 148 languages using improved language identification, data curation, mining, and teacher-student encoder training.The language identification system covers over 200 languages, while the encoder methodology enables transfer with minimal supervised bitext.Source: §5.4 · p. 46
• The authors release code and the stopes library for configurable, local or cluster-scale language identification, filtering, and bitext mining with open-source encoders.The tools support LASER encoders and other Huggingface sentence-transformers encoders, enabling broader high-quality bitext creation.Source: §5.4 · p. 46

6.2 Conditional Compute for Massively Multilingual Machine Translation

• Conditional computation uses sparsely gated MoE layers to increase capacity while limiting interference between unrelated language directions and preserving computational efficiency.MoE layers activate only a subset of parameters per input, routing each token to its top two experts.Source: §6 · p. 47, §6.2 · p. 50

6.1.1 Task Setup

• The multilingual setup uses a shared 256,000-token SentencePiece vocabulary with temperature-based sampling to improve low-resource language representation.The model is a Transformer encoder-decoder, and source-language prefixing improves zero-shot generalization to unseen source-target pairs.Source: §6.1.1 · pp. 48–49

6.1.2 Ablation Dataset

• The ablation benchmark covers 53 of 202 languages across 110 directions, with 864M examples spanning 40K to 186M examples per language pair.It includes 40 high-resource and 70 low-resource directions, including 22 very low-resource directions with fewer than 100K training examples.Source: §6.1.2 · pp. 49–50, Figure 15 · p. 49

6.2.1 Vanilla Sparsely Gated MoE and its drawbacks for Low-Resource Languages

• MoE-64 improves over dense 615M models by more than 1 chrF++ point on most direction subsets, but does not outperform dense 1.3B models at comparable compute.Very low-resource and non-English directions are exceptions to the improvements over dense 615M models.Source: Table 15 · p. 52, §6.2.1 · p. 53
• Overall dropout substantially improves MoE performance, including +1.5 to +1.7 chrF++ over dense 615M models across all subsets for the 615M compute-equivalent variant.For very low-resource directions translating out of English, dropout changes a relative decline of -0.1 chrF++ into a +0.9 chrF++ improvement.Source: Table 15 · p. 52, §6.2.1 · p. 53
• MoE models continue overfitting very low-resource directions because large capacity repeatedly processes small datasets, motivating architectural and training strategies that protect these pairs.The section evaluates these strategies on a benchmark designed to represent the final 200-plus-language setting.Source: §6 · p. 47, §6.2.1 · pp. 53–54

6.2.2 Regularizing Massively Multilingual Mixtures of Experts

• Expert Output Masking and Conditional MoE Routing outperform baseline overall dropout, with Table 16 finding EOM and CMR perform similarly.EOM masks selected expert outputs, whereas CMR learns whether tokens use a shared dense FFN or the higher-capacity MoE branch under a budget constraint.Source: §6.2.2 · pp. 54–56, Table 16 · p. 57
• +0.4 chrF++ across all pairs into English and +0.6 chrF++ across non-English pairs were obtained by MoE EOM over vanilla MoE with overall dropout, with +0.6 and +0.9 chrF++ gains out of English for low- and very low-resource languages.EOM also provided better regularization and less overfitting on low-resource tasks than overall dropout.Source: §6.2.2 · p. 57
• EOM outperformed FOM because masking before combining expert outputs strengthens the residual connection and reduces co-adaptation among experts and subsequent layers; CMR also improved over vanilla MoE but costs 18-25% more training compute.The authors therefore selected the simpler EOM strategy.Source: §6.2.2 · p. 57

6.2.3 Curriculum Learning

• Curriculum learning improved vanilla MoE by +0.6 chrF++ on low-resource and +0.8 chrF++ on very low-resource eng_Latn-xx directions, but did not help high-resource or xx-eng_Latn directions.Curriculum learning hurt MoE EOM on the ablation dataset, although the full dataset and larger model benefited from combining curriculum learning with EOM, especially for low- and very low-resource pairs.Source: §6.2.3 · p. 58, Table 17 · p. 59

6.2.4 Analysis of Multilingual Sparsely Gated MoE Models.

• MoE routing becomes language-specific in late decoder layers, with related languages choosing similar experts, whereas late encoder and early decoder layers are largely language-agnostic.UMAP embeddings show especially clear language-family separation in the decoder’s last layer, including Atlantic-Congo and Dravidian languages.Source: §6.2.4 · p. 59, Figure 20 · p. 60, Figure 21 · p. 61

6.3.3 Effect of Curriculum of Self-Supervision combined with Multilingual Machine Translation

• +1.3 and +1.9 chrF++ improvements on low- and very low-resource xx-eng_Latn pairs, respectively, were achieved by jointly training DAE and MMT, while DAE+MMT⇒MMT performed on par with joint training.DAE+MMT also improved low- and very low-resource eng_Latn-xx pairs by +0.4 and +1.1 chrF++, respectively, and improved xx-yy pairs by +1.1 chrF++.Source: §6.3.3 · p. 63, Table 18 · p. 63

6.3.4 Effect of Self-Supervision Objectives

• DAE was the most effective self-supervised objective with MMT: LM reduced performance, while combining DAE and LM also declined relative to DAE alone.The authors attribute LM’s weakness partly to lacking encoder-decoder attention training and suggest interference between DAE and LM when combined.Source: §6.3.4 · pp. 63–64, Table 19 · p. 64

6.4.1 Different sources of data

• The data-augmentation experiments distinguish primary human-translated data, mined data, neural MmtBT, and statistical SmtBT, while data tagging marks these secondary sources either separately, jointly, or not at all.Tagged backtranslation is intended to help models distinguish noisy data sources and avoid overfitting; the supplied passages describe the ablation setup but report no tagging result.Source: §6.4.1 · p. 65, §6.4.2 · p. 67
• The highest performance came from combining Primary, Mined, MmtBT, and SmtBT data, with SMT backtranslation still improving very low-resource directions.Although the MMT teacher generally outperformed SMT, the authors hypothesize that complementary noise sources explain SmtBT’s benefit.Source: §6.4.1 · pp. 66–67, Table 21 · p. 66

6.4.2 Data Tagging

• Finegrained tags produce the best chrF++ performance among no-tag, single-tag, and finegrained-tagging schemes, helping distinguish synthetic from natural data and signaling synthetic-data type.The comparison uses the Flores-200 ablation dev set and aggregates results by language direction and resource level.Source: §6.4.2 · p. 67, Table 23 · p. 68

6.5 Bootstrapping models with NLLB-Seed

• NLLB-Seed tests whether a few thousand professionally translated, broad-domain, language-verified sentences can bootstrap or adapt low-resource translation systems.The study compares PublicBitext, NLLB-Seed, their combination, and equal-sized Diminished+NLLB-Seed datasets.Source: §6.5 · pp. 67–68, §6.5.1 · p. 68, Figure 22 · p. 68

6.5.1 Usefulness of NLLB-Seed

• NLLB-Seed training outperforms substantially larger PublicBitext data on average, while combining both datasets performs best and replacing equal-sized public data with NLLB-Seed improves results.Repeated experiments show small variance, supporting lower PublicBitext performance as a quality issue rather than sampling bias.Source: §6.5.1 · p. 69, Table 24 · p. 69, Table 25 · p. 70

6.5.2 Effect of NLLB-Seed on Backtranslation

• Backtranslation increases the NLLB-Seed advantage: it yields +1.6 chrF++ from PublicBitext versus +3.5 chrF++ from Diminished+NLLB-Seed data.The compared datasets have the same total size, indicating that high-quality bitext improves model-based data augmentation.Source: §6.5.2 · p. 70, Table 25 · p. 70

6.6 Human Evaluation

• Human evaluation indicates that +0.5 chrF++ is often detectable, while the best MoE, SSL, and BT variants produce statistically significant gains across multiple directions.For the MoE, SSL, and BT variants, significant improvements occurred in 10, 4, and 10 directions, respectively.Source: §6.6 · p. 71

6.7 Conclusion

• MoE capacity, regularization, and curriculum training improve low-resource directions while maintaining high-resource performance, while self-supervision and diverse backtranslation address data scarcity.The methods reduce interference between unrelated language directions and combine backtranslated data from MMT and SMT models with mined data.Source: §6.7 · p. 71

7.1 Automatic Evaluation

• The evaluation uses BLEU and chrF++, primarily relying on chrF++ because character n-grams handle languages without word spaces and agglutinative morphology.For comparisons with published work, the study uses BLEU and spBLEU where appropriate; spBLEU is enabled for 200+ languages through the released SPM-200 model.Source: §7.1 · pp. 72–73

7.2.1 Methodology

• XSTS prioritizes meaning preservation over fluency by rating source–translation equivalence on a five-point scale, reflecting the goal of usable translations for low-resource languages.The protocol is adapted from crosslingual Semantic Text Similarity and has higher interannotator agreement than Direct Assessment in cited work.Source: §7.2.1 · pp. 73–74
• Calibrated XSTS scores improve correlation with automated metrics, while the explored calibration methods show no dramatic differences from one another.Evaluators rate a shared English calibration set to identify systematic harshness or generosity, after which raw task scores are adjusted; the calibration set uses 1,000 backtranslated Flores-200 sentence pairs.Source: §7.2.1 · pp. 74–76

7.2.2 Results

• Human evaluation covers five translation models and 86 language directions, with up to 292 evaluators rating Flores-200 translations.Each source–translation pair was generally scored by three evaluators, using 1,000 Flores-200 sentences or randomly selected 500-sentence subsets.Source: §7.2.2 · p. 77
• 0.710, 0.687, and 0.694 are the Spearman R correlations of aggregated XSTS with spBLEU, corpus chrF++, and sentence-level chrF++, respectively.Corpus spBLEU has the strongest nominal correlation, followed by sentence-level chrF++ and corpus chrF++; calibrated human scores correlate more strongly with automated metrics than uncalibrated scores.Source: Figure 24 · p. 76, §7.2.2 · p. 77
• Human score distributions show high-quality, variable-quality, and predominantly poor or incoherent translation patterns across language directions.The variable-quality category can have a high average despite many poorly rated sentences, whereas the poorest category contains many XSTS scores of 1 and incoherent outputs.Source: §7.2.2 · p. 78, Figure 25 · p. 78
• Into-English translation quality is typically higher than out-of-English quality in both automated and human evaluations, with exceptions including snd and azj.The comparison uses chrF++ and mean median XSTS scores for translations generated by M2M-100 (12B).Source: §7.2.2 · p. 78, Figure 26 · p. 79

7.3 Toxicity

• The toxicity analysis releases wordlists for 200 languages and targets added toxicity, which can arise from mistranslation or hallucination and undermine user trust.The work also proposes scalable detectors and filtering strategies, while defining toxicity to include vulgar or profane language, sexual content, hate speech, and bullying expressions.Source: §7.3 · p. 79, §7.3.2 · p. 80

7.3.2 Toxicity Lists for 200 Languages

• The toxicity lists cover all 200 languages, with a median of 143 entries, a mean of 271, and lengths ranging from 36 to 6,078 entries.Lists primarily include profanities, insults, pornographic terms, hate speech, bullying terms, and sexually associated body-part terms.Source: §7.3.2 · p. 81

7.3.3 Toxicity Detection

• 66.8% F1 in English and 61.6% F1 in French exceed better-than-chance baselines, while detectors remain better than chance on low-resource languages.Performance declines for 2+ toxic-item detectors and is lowest in several languages partly because their scripts are inadequately tokenized.Source: §7.3.3 · p. 83, Table 26 · p. 83
• Toxicity varies across corpora and languages, and sampled mined-data toxicity often reflects misaligned bitext with omissions or detoxification that can encourage toxic mistranslations.Filtering sentences with imbalanced multiple toxic items improves translation performance and tends to reduce toxic items in outputs.Source: §7.3.3 · pp. 83–85

7.3.4 Open Challenges in Toxicity for Machine Translation

• Added toxicity can arise from hallucination or mistranslation, and ALTI+ distinguishes these cases by comparing source and target-prefix contributions.Low source contribution indicates hallucination in the Wolof-to-English and Quechua-to-English examples, whereas high source contribution identifies mistranslation or accurate toxic translation.Source: §7.3.4 · p. 86, §7.3.3 · p. 86, Figure 31 · p. 87

7.4 Conclusion

• Evaluation uses spBLEU and chrF++ for multilingual automatic scoring, then XSTS and calibration to make human scores comparable across language pairs.The work detects and mitigates added toxicity across all 200 languages using toxic-term lists and classifiers, while acknowledging cultural and linguistic biases and allowing translators to reject toxic-language work.Source: §7.4 · p. 89, §7.3.5 · p. 89

8. Bringing it All Together

• The project open-sources its evaluation benchmarks, models, datasets, training and inference code, and other modeling and data artifacts.The work combines interdisciplinary methods to improve multilingual translation for underserved low-resource communities and support community scrutiny and reuse.Source: §8 · p. 90

8.1.4 Filtering Strategy

• Filtering improved translation performance while reducing hallucinated toxicity on experiments covering six translation directions.The pipeline included checks for mining quality, sentence lengths, language identity, toxicity imbalance, and duplicate data.Source: §8.1.4 · p. 95

8.1.5 Effect of using Different Data Sources on Performance

• Mining and backtranslation significantly improved performance across language-pair types and resource levels, with larger gains for low-resource than high-resource languages.Adding Mined data improved over Primary data, and adding MmtBT+SmtBT produced further improvements.Source: §8.1.5 · p. 95, Figure 34 · p. 95
• +18.6 chrF++ overall improvement was achieved for very low-resource into-English directions by adding mined data and backtranslation.Mined data contributed +12.5 chrF++, while additional backtranslated data contributed +6.1 chrF++ for languages with fewer than 100K public-bitext sentence pairs.Source: §8.1.5 · p. 96
• +8.9 chrF++ overall improvement was achieved for very low-resource non-English-centric pairs through mining and backtranslation.Out-of-English directions improved by +6.6 chrF++ overall, including +4.7 chrF++ from mined data and +1.9 chrF++ from backtranslated data.Source: §8.1.5 · p. 96

8.1.6 The 200 Language Dataset

• The final training dataset covers 202 languages with 1,220 language pairs, 2,440 directions, and over 18B sentence pairs.It combines primary bitext for 661 pairs, mined bitext for 784 pairs, and 261 backtranslated directions; arb_Latn and min_Arab lack training data.Source: §8.1.6 · p. 96

8.2.1 Does Self-supervised Learning help on top of Mining and Backtranslation?

• Self-supervised learning gains diminished as mined and backtranslated data were added and became negligible with both backtranslation sources.Because MMT+SSL performed worse on eng_Latn-xx directions after backtranslation, the final model uses only the MMT task.Source: §8.2.1 · p. 98, Figure 36 · p. 98

8.2 Preparing the Model

• NLLB-200 combines a 3.3B-parameter dense baseline, sparsely gated Mixture-of-Experts capacity, regularization, and curriculum learning to handle 2,440 directions.MoE increases capacity while limiting proportional FLOP growth, but regularization and curriculum learning address rapid low-resource overfitting and multilingual interference.Source: §8.2 · p. 97, §8.2.2 · p. 99

8.2.2 Scaling Model Architecture

• Every fourth Transformer block uses a 128-expert MoE layer, while EOM regularization is selected because it uses 23% fewer FLOPs than CMR for a favorable accuracy-efficiency trade-off.Among insertion frequencies, fMoE=4 performs best and is 28% more efficient than fMoE=2.Source: §8.2.2 · pp. 99–100, Table 28 · p. 100, §8.2.4 · pp. 101–102

8.2.3 Designing an Optimized Training Curriculum

• The four-phase curriculum outperforms no curriculum, improving xx-eng_Latn by +1 chrF++ overall and by +1.9 chrF++ on very low-resource directions.Compared with naive bucketing, phased curriculum improves low-resource and very low-resource xx-eng_Latn directions by +0.7 and +1.2 chrF++, respectively.Source: §8.2.3 · p. 101, Table 29 · p. 101

8.2.4 The 200 Language Model: NLLB-200

• The final NLLB-200 is a 54.5B-parameter Transformer with 24 encoder and decoder layers, 128 experts, and FLOPs similar to a 3.3B dense model.It is trained for 300k steps with the four-phase curriculum, a 1M-token effective batch, and EOM dropout with p_eom=0.2.Source: §8.2.4 · pp. 101–102

8.3.1 Performance on Flores-101 and Comparison to State-of-the-Art

• +7.3 spBLEU on average, a 44% improvement over the nearest state-of-the-art on Flores-101, despite covering over 200 languages and more than 30k additional directions.On the overlapping 87-language subset, NLLB-200 also improves by +6.9 spBLEU on average.Source: §8.3.1 · pp. 102–103, Table 30 · p. 103

8.3.2 Performance on Flores-200

• On Flores-200, NLLB-200 averages 35.6 chrF++ across 40,200 non-English-centric directions and 35.4 chrF++ on 38,162 zero-shot directions.Zero-shot performance is 4.1 chrF++ below approximately 2,000 supervised directions, while high-resource English-out directions exceed low-resource ones by 13.0 chrF++ on average.Source: §8.3.2 · pp. 105–106

8.3.3 Human Evaluation

• NLLB-200 achieves an average XSTS score of 4.22 versus 3.66 for the same-data 3.3B dense baseline, a statistically significant improvement across 51 directions.Scores average 4.09 into English, 4.33 out of English, and 4.12 for non-English directions.Source: Figure 37 · p. 107, §8.3.3 · p. 107

8.3.4 Prevalence of Toxicity

• Potential added toxicity occurs in 0.80% of outputs translating out of English versus 0.27% into English on Flores-200, with low overall prevalence but domain-dependent uncertainty.The evaluation detects potential added toxicity in 130 of 201 out-of-English directions and 200 of 201 into-English directions, and warns that other domains may differ.Source: §8.3.4 · p. 108

8.4 Out-of-domain Generalization: Performance on non-Flores-200 Domains

• The study next evaluates NLLB-200’s generalization capability on non-Flores machine-translation benchmarks.This subsection introduces out-of-domain evaluation, but the supplied passage does not report its results.Source: §8.4 · p. 108

8.4.1 Public Benchmarks

• Across FLORES(v1), WAT, WMT, TICO, MAFAND-MT, Autshumato, and MADAR, evaluation covers 238 directions using benchmark-matched BLEU, spBLEU, and chrF++ procedures.The benchmark suite spans news, health, governmental, conversational, medical, African-language, and Arabic-dialect data, with comparisons to previous state-of-the-art systems.Source: §8.4.1 · pp. 108–110

8.4.2 Results

• NLLB-200 outperforms prior state of the art on most evaluated out-of-domain benchmarks, including 21 of 24 IWSLT directions and 14 of 26 MAFAND-MT directions.It improves IWSLT translation from English by 3.5 BLEU points and into English by 6.3, while MAFAND-MT improves into English by 4.0 BLEU points but declines in several other directions.Source: Table 36 · p. 110, §8.4.2 · p. 111, Table 38 · p. 113

8.4.3 Effective Domain Adaptation with NLLB-MD

• Finetuning NLLB-200 on only 2,000 high-quality NLLB-MD sentences improves chrF++ by 7.7 in chat, 3.1 in news, 4.1 in health, and 5.8 in scripted domains.The setup uses 500 held-out test sentences per language and single-task finetuning in both directions; the best strategy omits load-balancing regularization.Source: §8.4.3 · pp. 112–113, Figure 39 · p. 114
• Without load balancing, finetuning improves average chat-domain accuracy by +7.2 chrF++ points, compared with +6.1 points with load balancing.The no-load-balancing strategy particularly benefits the five low-resource target languages, increasing accuracy by +3.7 chrF++ points.Source: §8.4.3 · p. 112, Figure 38 · p. 113

8.5.1 Language Co-location in NLLB-200 experts

• NLLB-200 experts co-locate languages by family and geographic or genealogical relatedness, with early encoder and late decoder layers dispatching related languages to similar expert sets.The analysis computes per-language distributions across 128 experts on the FLORES-200 development set and embeds them with UMAP.Source: §8.5.1 · p. 114, Figure 40 · p. 115

8.5.2 Effect of Phased Curriculum on Low-Resource Overfitting

• The phased curriculum mitigates low-resource overfitting more effectively than naive curricula, especially for very low-resource pairs, though it remains non-optimal.Introducing pairs too early or too late can hurt performance, while high-resource pairs are affected little; automated curricula are left for future work.Source: Figure 41 · p. 115, §8.5.2 · p. 116

8.5.3 Impact of Multilingual Transfer

• Multilingual transfer produces gains over bilingual models on almost all compared low-resource directions, with more consistent improvements into English and direction-specific changes elsewhere.English-to-low-resource directions include gains as large as +19.4 chrF++ and decreases as small as -0.6 chrF++.Source: §8.5.3 · p. 116, Table 39 · p. 116

8.6 Making Large Models More Accessible through Distillation

• Distillation compresses NLLB-200 into lower-latency models for Wikipedia Content Translation, supporting 25 languages and 74 directions, while also targeting compute-efficient distillation across all 200 languages.The work investigates word-level online distillation and sentence-level offline distillation, using a fine-tuned 1.3B dense teacher for Wikipedia-domain models.Source: §8.6 · p. 117, §8.6.1 · p. 117, §8.6.2 · pp. 117–118

8.6.2 Creating Models Specialized for the Wikipedia Domain

• Offline distillation marginally outperforms online distillation for Wikipedia specialization, with the 515M offline student improving over the 1.3B teacher by +0.4, +0.4, and +0.3 chrF++ on English, French, and Spanish directions.Online distillation remains compute-efficient because it avoids autoregressively generating millions of translations with a very large MoE teacher.Source: §8.6.2 · pp. 118–119

8.6.3 Distillation of NLLB-200, a 54B Parameter MoE Model

• +0.5 chrF++ and +0.3 chrF++ are the average gains from distillation over dense 1.3B and 615M baselines, respectively, although students remain -1.4 and -3.7 chrF++ below NLLB-200.For the 1.3B setting, the distilled model gains +0.6 chrF++ over its baseline, with larger improvements for low- and very low-resource languages.Source: §8.6.3 · pp. 119–120, Table 41 · p. 120

8.6.4 Conclusion.

• Distillation is presented as a promising way to specialize multilingual models by language directions and domain, enabling strong Wikipedia translation despite a 1.3B-parameter teacher.The authors open-source distilled models and an offline distillation pipeline to facilitate access and reuse.Source: §8.6.2 · p. 118, §8.6.4 · p. 120

8.7.1 Transliteration

• Specialized transliteration models perform best, but error rates remain relatively high across approaches, underscoring the need to reduce script-based disparities.Ejawi improves substantially with rule-based character replacement, whereas neural systems could plausibly benefit from larger training datasets.Source: §8.7.1 · pp. 121–122, Table 42 · p. 122

8.7.2 Multidialectal Translation

• Arabic languoid translation quality tracks similarity to MSA: Najdi is closest and performs best, whereas Tunisian and Mesopotamian are more divergent and perform worse.For generated English-to-Arabic text, sentence-level dialectness standard deviation decreases from 14.6 in evaluation data to 8.8.Source: §8.7.2 · pp. 124–126
• NLLB-200 outperforms a smaller Arabic-focused model on average and leads in 17/28 directions, but the smaller model remains better when translating into Arabic languoids.NLLB-200 benefits from over 53M unique primary French sentences versus under 38M for the custom model, while focused modeling remains advantageous for some directions.Source: §8.7.2 · pp. 126–127

8.8 Environmental Impact of NLLB

• NLLB-200 training used 51,968 A100 GPU-hours and was estimated at 8.39 tCO2eq, fully offset by the provider’s sustainability program, while the estimate excludes several power and manufacturing costs.The authors caution that carbon estimates are imprecise because GPU utilization, non-GPU infrastructure, cooling, and manufacturing emissions are not fully measured.Source: §8.8 · p. 127
• The carbon assessment reports GPU-using stages throughout the research process instead of estimating total emissions from final-model training alone.This approach provides insight into the compute requirements and efficiency of each stage.Source: §8.8 · p. 127, Table 48 · p. 128
• Sparse activation can make Mixture-of-Expert models more compute-efficient than dense counterparts, while careful workload design can lower power consumption and emissions.The paper urges researchers to measure environmental impact, optimize energy efficiency, and open source materials to avoid duplicated workloads.Source: §8.8 · pp. 127–128
• 104.31 tCO2eq were estimated for the entire NLLB project, with 100% of emissions directly offset by the provider.The accounting covers data preparation, modeling experiments, final-model ablations, and evaluations; experimentation was the most compute-intensive phase.Source: §8.8 · p. 127, Table 48 · p. 128

9. No Language Left Behind: Social Impact & Concluding Thoughts

• NLLB contributes Flores-200, NLLB-Seed, low-resource bitext mining, overfitting-reduction methods, and distilled models for more accessible deployment.NLLB-Seed contains human-translated bitext for 43 languages, while the mining approach creates hundreds of millions of aligned training sentences.Source: §9 · p. 128

9.1 Expanding Information Access

• NLLB could expand low-resource communities’ access to books, research, archives, news, social media, and encyclopedias, while improving web accessibility and knowledge acquisition for existing users.Better translation could also reduce the cognitive effort required to decipher poorly translated content.Source: §9.1 · p. 129

9.2 The Janus-faced Nature of Digital Participation

• NLLB could support localized cultural knowledge and minority-language revitalization, but greater digital participation may also increase exposure to surveillance, censorship, misinformation, scams, and hate speech.The authors propose strengthening cross-lingual detection tools while pairing new technology with structural investments and improved online literacy.Source: §9.1 · p. 129, §9.2 · p. 130
• NLLB may deepen digital inequities because technological skills, demographic conditions, and uneven infrastructure determine who benefits from expanded translation access.The authors stress reflexive expansion, relationships with local institutions and communities, and collaboration with invested research groups.Source: §9.2 · p. 130

9.3 The Future of NLLB: A Collective Responsibility

• Open-sourcing NLLB promotes transparency, reduces duplicated effort, and enables collaboration, but equitable use still requires support for under-resourced researchers and diverse interdisciplinary participation.Planned toolkits and grants aim to address computing disparities, while future teams should include underrepresented humanities, social-science, racial, gender, and cultural perspectives.Source: §9.3 · pp. 130–131
• Making the web accessible across languages requires more than NLLB: education, internet access, digital literacy, and other policy interventions must address structural language disparities.The authors frame NLLB as one component of a multifaceted, stakeholder-driven effort.Source: §9.3 · p. 131

10. Contributions

• The contributions section recognizes that many individuals supported multiple areas and groups contributors by section and alphabetically by last name.It notes that each person is mentioned once despite contributing across several areas.Source: §10 · p. 131

Data

• Data contributors built monolingual cleaning, language-identification, mining, bitext, and data-partnership infrastructure, including LASER3 and open-sourced mined data.These efforts covered CommonCrawl processing, deduplication, filtering, language identification for 200+ languages, and mined bitext production.Source: Data · pp. 131–132, data · p. 132

data

• Additional data contributors led research and engineering for mining, LASER3, bitext development, data quality, and Wikipedia model deployment.Their roles included technical feedback, data open-sourcing, infrastructure implementation, and research leadership for data development.Source: data · p. 132, Data · p. 132

Modeling

• Modeling contributors advanced mixture-of-experts architectures, backtranslation, self-supervised learning, regularization, distillation, curriculum learning, and model-quality analysis.They also investigated NLLB-Seed, transliteration, data quality, Wikipedia translation models, and the properties of MoE models.Source: Modeling · p. 132

Evaluation

• Evaluation contributors collected and annotated FLORES 200, NLLB-Seed, and human-evaluation data while studying translation quality, toxicity, ethics, and human impacts.Their work included XSTS, interpretability and mitigation of toxicity, interview studies, human-evaluation experiments, and research artifacts such as data and model sheets.Source: Evaluation · pp. 132–133

Linguistics

• Linguistics contributors addressed Arabic dialectal variation, transliteration, language identification, FLORES 200 language metadata and quality, output quality, toxicity detection, and ethics.Their contributions supported linguistic resource development and analysis of model behavior across languages.Source: Linguistics · p. 133

Organization

• Project leaders provided research, engineering, evaluation, data, program, product, and open-source direction, including collaboration with the Wikimedia Foundation.The organization covered overall strategy, research direction, management, technical program coordination, and product leadership.Source: Organization · p. 133

Ethical framing

• The paper examines centralized and distributedly standardized languages because language technologies can affect language use and evolution.This framing connects model development to responsibility toward language communities.Source: §A.1 · p. 167

Centralized standardization

• For centrally standardized languages, institutional grammars, dictionaries, and neologism policies can guide data collection, annotation, and dataset filtering, while web-mined data captures unsanctioned usage.Web data may reflect more fluid language development than official standards.Source: §A.1 · p. 167

Distributed standardization

• For distributedly standardized languages, conflicting or weak standards make annotation guidelines potentially favor one variant, leaving communities feeling excluded or pressured to adopt alien standards.Many such languages are low-resource because annotated data is absent or extremely scarce.Source: §A.1 · p. 167
• Closely related variants can benefit from transfer learning, while transliteration and spelling-standard mapping can extend support with limited additional annotation.These approaches may allow competing variants to benefit from one another’s presence in the dataset.Source: §A.1 · pp. 167–168

B.1 Toxicity Lists

• Human translation created non-English toxicity lists from English prompts, with German- and Spanish-mediated pivots for Luxembourgish and Asturian, followed by iterative prompt revisions to reduce ambiguity-driven errors.The prompt was modified during the process as translators encountered misunderstandings and translation errors.Source: §B.1 · p. 168
• For most FLORES 101 languages, existing lists allowed broad additions of cultural specificities, spelling variants, inflections, leetspeak, and nonstandard forms.Because many entries lacked identifiable dictionary forms, quality and anomaly checking was difficult, while string matching could not exhaustively cover nonstandard variants.Source: §B.1 · p. 168
• For newly added languages, prompts supplied part-of-speech, register, and definition links, while translators could identify missing equivalents and suggest roughly forty frequent word forms.This method did not request leetspeak or nonstandard variants, though translators could include them voluntarily.Source: §B.1 · p. 168
• Both list-construction methods struggled to distinguish slurs from nontoxic language, especially for racial and LGBTQ+ terms, despite richer prompts and complementary training.The difficulty remained salient even when additional grammatical and semantic information was provided.Source: §B.1 · p. 168
• Judgments were further complicated by limited access to translators with diverse backgrounds and by terms potentially reclaimed by the communities they target.Even non-member allies may hesitate to use reclaimed terms, making classification decisions difficult.Source: §B.1 · p. 169

Appendix C. Data

• The LID system is evaluated with precision, recall, and F1 across all Flores-200 languages.Results are reported on the Flores-200 devtest set in Table 49.Source: Appendix C · p. 170, Table 49 · p. 172
• A domain mismatch exists between the data used to train the LID system and the web corpus used in the pipeline.The mismatch motivated an extensive human evaluation.Source: Appendix C · p. 170
• Human annotators assessed several thousand randomly selected sentences predicted to belong to 74 low-resource languages with low preliminary-model F1 scores.These annotations were used to determine whether each prediction was correct.Source: Appendix C · p. 170
• The human annotations were used to construct a challenge set for benchmarking the final LID model.Source: Appendix C · p. 170
• Table 49 presents LID results covering all languages in the Flores-200 dataset.Source: Table 49 · p. 172
• Table 50 compares the final LID model with CLD3 using F1 and false positive rates on the human-annotated challenge set.FPR is reported because F1 can be misleading under web-data class balances different from the development set; unsupported CLD3 languages have blank cells.Source: Appendix C · p. 170, Table 50 · p. 173

D.2 SMT vs MMT

• Combining SMT and MMT backtranslations is beneficial because the two models provide different, complementary sources of noise.Figure 42 visualizes token-frequency differences for kau_Arab-to-eng_Latn and awa_Deva-to-eng_Latn translations.Source: §D.2 · p. 174, Figure 42 · p. 174

D.1 Ablation Dataset

• The ablation dataset contains exactly 110 language pairs used for experiments in Section 6.Table 51 lists the languages and language pairs included in this experimental dataset.Source: §D.1 · p. 174, Table 51 · p. 175

E.1.2 Importance of Backtranslation Quality on Model Scaling

• Backtranslation, self-training, and large-scale mining supply substantial training data, but their lower quality than human translations may limit overall model quality.The section studies whether data-augmentation quality constrains model scaling.Source: §E.1.2 · p. 175
• The scaling experiment trains a multilingual model on 60 African-language directions to and from French and English, then evaluates backtranslation quality for eight languages.The evaluated languages are fuv_Latn, kmb_Latn, lug_Latn, nya_Latn, swh_Latn, umb_Latn, wol_Latn, and zul_Latn.Source: §E.1.2 · pp. 175, 177

E.1.1 Primary Dataset Composition

• NLLB-200 training uses datasets largely downloaded through OPUS and the mtdata tool, with direction counts excluding reverse directions.Complete training configuration files are available in the project repository.Source: §E.1.1 · p. 175, Table 52 · p. 176

E.1.2 Importance of Backtranslation Quality on Model Scaling

• Performance quickly plateaus when larger final models are trained on backtranslations generated by small models.The study varies backtranslation-generating models of 219M and 1.4B parameters and final models of 219M, 514M, 776M, and 1.4B parameters across eight African languages.Source: Figure 43 · p. 177, §E.1.2 · p. 177

E.1.3 Training Directions and Curriculum Buckets.

• Training directions are organized into Primary, Mined, Primary+Mined, and Primary+Mined+MmtBT+SmtBT data-source configurations, alongside a listed 4-phase curriculum.The appendix provides separate language-pair files for each training-data combination and curriculum setup.Source: §E.1.3 · p. 177

E.2.1 Performance on African Languages

• On Flores-101 non-English-centric directions, Table 53 compares NLLB-200 with MMTAfrica and reports that NLLB-200 outperforms on most translation directions.The table reports spBLEU/chrF++ and bolds the best score.Source: Table 53 · p. 178, §E.2.1 · p. 178

E.2.2 Comparison against Google Translate

• NLLB-200 performs slightly better overall than Google Translate, with significantly stronger results on low- and very low-resource directions, although Google Translate leads on some English-to-other high-resource pairs.The comparison covers 206 English-centric Flores-200 directions; professional machine-translation post-editing may advantage Google Translate on heavily used high-resource pairs.Source: §E.2.2 · pp. 178–179, Table 54 · p. 179

E.3 Toxicity Evaluation

• Toxicity detections are dominated by benign-term over-detection, but some detected cases reflect genuinely hallucinated toxicity, especially inconsistently across out-of-English directions.A small baseline detection level appears on nearly all into-English directions, whereas out-of-English detection varies across both very high- and very low-resource languages.Source: §E.3 · p. 179, Figure 44 · p. 180

E.4 Out-of-domain Generalization: Performance on non Flores-200 Domains; E.5 Analysis of NLLB-200; E.6 Full Distillation Results.

• The paper presents full evaluations on non-Flores-200 datasets and domains, analyzes language similarity across Flores-200, and reports complete Wikipedia-domain distillation results.The distillation evaluation uses offline sequence-level distillation for a smaller Wikipedia-domain model and includes English, French, Spanish, and additional requested directions.Source: §E.4 · p. 179, §E.5 · p. 179, §E.6 · p. 179

Appendix F. Model Card - NLLB-200; Intended Use

• NLLB-200 is a Transformer Mixture-of-Experts research model developed by Meta AI Research, released as version NLLB-200 on June 30, 2022.The model card identifies research in machine translation, especially for low-resource languages, as its primary intended use.Source: Appendix F · p. 185, Intended Use · p. 185

Intended Use

• NLLB-200 supports single-sentence translation among 200 languages but is not intended for production, domain-specific or document translation, certified translations, or inputs longer than 512 tokens.The model card also warns that mistranslations, misinformation, scams, residual personally identifiable information, and harms in health or safety decisions remain possible.Source: Intended Use · pp. 185–186

Caveats and Recommendations; Carbon Footprint Details

• The model has limited investigation beyond the Wikimedia domain, may miss language variations, and reports its CO2e estimate in Section 8.8.Users are advised to make appropriate assessments when applying the model, while the carbon-footprint details point to the paper’s dedicated estimate.Source: Caveats and Recommendations · p. 186, Carbon Footprint Details · p. 186

Dataset Descriptiona

• NLLB-Seed contains 6,193 human-translated English sentences covering 43 language pairs and 11 knowledge categories, providing clean starter data for low-resource translation.The source topics include anthropology, arts, biology, geography, history, mathematics, people, philosophy, physical sciences, society, and technology.Source: Dataset Descriptiona · p. 187
• The dataset was created to boost low-resource translation performance and improve language identification for extremely low-resourced languages.The data is accessible through the project README.Source: Dataset Descriptiona · p. 187, Considerations for Using the Data · p. 187
• Curation aligned scripts, dialects, spelling, and translation approaches with Flores-200, using direct English translation except for specified Ligurian and transliteration workflows.Completed translations underwent linguistic assessment against Flores-200 alignments and basic quality sanity checks.Source: Dataset Descriptiona · p. 187

Considerations for Using the Data

• NLLB-Seed has no extra annotations, and its dataset biases have not been studied.The data card reports no applicable personal or sensitive information.Source: Dataset Descriptiona · p. 187, Considerations for Using the Data · p. 187
• Translators were vetted as native target-language speakers with advanced English proficiency, relevant experience, and periodic translation testing; the Wikipedia-based translations use CC-BY-SA.Non-English translators also required high fluency in their source language, with experience requirements varying by academic qualification.Source: Considerations for Using the Data · p. 188
• The data card identifies the accompanying paper as NLLB Team et al., No Language Left Behind: Scaling Human-Centered Machine Translation, Arxiv, 2022.The card follows a documented dataset-card template.Source: Considerations for Using the Data · p. 188

Dataset Summary

• 11,810 human-translated sentences span news, formal speech, informal speech, and medical sources for six English language pairs targeting low-resource languages.The languages are Central Aymara, Bhojpuri, Dyula, Friulian, Russian, and Wolof.Source: Appendix H · p. 189

Curation Rationale

• The data was translated directly from English after Flores-200 established script, dialect, spelling, and translation alignments, followed by linguistic and basic quality checks.Completed datasets were checked against the Flores-200 alignments before finalization.Source: Appendix H · p. 189

Source Data

• Source material comprises 2,810 WMT21 English-German development sentences for news, 3,000 multi-session chat utterances for informal speech, and 3,000 health-report sentences.The health sources include a World Health Organisation report and the English portion of the TAUS Corona Crisis Report.Source: Appendix H · p. 189

Considerations for Using the Data

• The dataset is intended to improve translation quality and language identification for extremely low-resource languages, but its biases have not been studied.It is also used to boost NLLB performance on low-resource languages and assess fine-tuning on new domains.Source: Appendix H · p. 189

Dataset Curators

• Translators were vetted native speakers with high English proficiency and relevant experience, and source-based translations were released under CC-BY-SA or CC-BY-NC licenses.Translators also undergo periodic translation testing; no extra annotations or personal and sensitive information are included.Source: Appendix H · pp. 189–190

Dataset Summary

• 148 English-centric and 1,465 non-English-centric language pairs are represented in mined bitext reconstructed with stopes and LASER3 encoders.The metadata and reconstruction script are open sourced, using specified CommonCrawl and ParaCrawl files.Source: Appendix I · p. 191

Metadata Format

• Metadata files are compressed, space-separated records containing source locations, digests, language-identification and LASER scores, directions, and line indices.Common Crawl records use web-file and document identifiers, while other-corpus records use corpus and language identifiers; digests use xxh3_64_intdigest.Source: Appendix I · p. 191

Limitations and Risks

• Because the mining process is noisy, the data card recommends using the reconstructed bitext for training and Flores-200 or other datasets for evaluation.Common Crawl and ParaCrawl data may also contain personally identifiable, sensitive, or toxic content publicly shared online.Source: Appendix I · pp. 191–192

Source Data and Curation

• The source data comes from Common Crawl and ParaCrawl, filtered using language identification, emoji removal, and language-model filtering for some high-resource languages.Parallel sentences were identified with LASER3 encoders, and the data was not human annotated.Source: Appendix I · pp. 191–192

Biases and Curation

• Internet-derived data likely contains prevalent web biases, while language identification and filtering may disproportionately affect lower-resource languages and remove less natural utterances.The biases were not specifically studied, and the dataset was not curated.Source: Appendix I · p. 192

Licensing Information

• The released metadata and reconstruction script use CC-BY-NC, while original text and copyright remain governed by source authors’ or publishers’ applicable licenses.Source URLs are retained for sentence-level licensing reference.Source: Appendix I · p. 192