{"id":3236,"date":"2026-11-04T10:47:01","date_gmt":"2026-11-04T05:17:01","guid":{"rendered":"https:\/\/www.ulatus.com\/translation-blog\/?p=3236"},"modified":"2026-10-09T10:49:18","modified_gmt":"2026-10-09T05:19:18","slug":"why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it","status":"publish","type":"post","link":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/","title":{"rendered":"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It"},"content":{"rendered":"<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_84 counter-flat ez-toc-counter ez-toc-custom ez-toc-container-direction\">\n<div class=\"ez-toc-title-container\">\n<span class=\"ez-toc-title-toggle\"><\/span><\/div>\n<nav><ul class='ez-toc-list ez-toc-list-level-1 ' ><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#The_Data_Wall_Hits_Non-English_First\" >The Data Wall Hits Non-English First<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#What_Is_the_Multilingual_Training_Data_Bottleneck\" >What Is the Multilingual Training Data Bottleneck?<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#The_four_forms_it_takes\" >The four forms it takes<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Why_Is_Multilingual_Training_Data_the_Bottleneck_for_LLMs_in_2026\" >Why Is Multilingual Training Data the Bottleneck for LLMs in 2026?<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Skewed_web_data_and_the_long_tail\" >Skewed web data and the long tail<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#The_tokenization_tax\" >The tokenization tax<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Alignment_data_scarcity\" >Alignment data scarcity<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Why_compute_alone_does_not_fix_it\" >Why compute alone does not fix it<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-9\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#How_Do_You_Diagnose_Which_Data_Bottleneck_You_Actually_Have\" >How Do You Diagnose Which Data Bottleneck You Actually Have?<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-10\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Quick_tests_worth_running_this_week\" >Quick tests worth running this week<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-11\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Why_diagnosing_before_buying_saves_budget\" >Why diagnosing before buying saves budget<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-12\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#How_Can_LLM_Builders_Overcome_Multilingual_Data_Shortages\" >How Can LLM Builders Overcome Multilingual Data Shortages?<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-13\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#1_Fix_volume_targeted_collection_community_corpora_and_consent\" >1. Fix volume: targeted collection, community corpora and consent<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-14\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#2_Fix_quality_language-ID_filtering_deduplication_and_native_review\" >2. Fix quality: language-ID filtering, deduplication and native review<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-15\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#3_Fix_alignment_native-written_not_just_translated\" >3. Fix alignment: native-written, not just translated<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-16\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#4_Fix_evaluation_per-language_culturally_grounded_test_sets\" >4. Fix evaluation: per-language, culturally grounded test sets<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-17\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Guardrail_for_all_four_translate_then_post-edit\" >Guardrail for all four: translate, then post-edit<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-18\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Build_Crowdsource_Synthesize_or_Partner_Which_Sourcing_Model_Fits\" >Build, Crowdsource, Synthesize or Partner: Which Sourcing Model Fits?<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-19\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Decision_rule_by_stage\" >Decision rule by stage<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-20\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#When_a_specialist_partner_wins\" >When a specialist partner wins<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-21\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#A_90-Day_Roadmap_to_Close_Your_Multilingual_Training_Data_Gap\" >A 90-Day Roadmap to Close Your Multilingual Training Data Gap<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-22\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Days_1-30_audit_and_baseline\" >Days 1-30: audit and baseline<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-23\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Days_31-60_pilot_fixes_on_two_priority_languages\" >Days 31-60: pilot fixes on two priority languages<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-24\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Days_61-90_scale_evaluate_and_set_governance\" >Days 61-90: scale, evaluate and set governance<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-25\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#Treat_Multilingual_Training_Data_as_a_Product_Not_a_Backlog_Item\" >Treat Multilingual Training Data as a Product, Not a Backlog Item<\/a><\/li><\/ul><\/nav><\/div>\n<h2><span class=\"ez-toc-section\" id=\"The_Data_Wall_Hits_Non-English_First\"><\/span>The Data Wall Hits Non-English First<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Ask the same 11,829 questions in 29 languages and Qwen2.5-72B scores 70.3% in English but only 40.1% in Swahili. That result comes from MMLU-ProX, a 2025 benchmark of 36 models that found gaps of up to 24.3% between high- and low-resource languages. And the pattern points to one place: multilingual training data.<\/p>\n<p>I call this the data wall: the point where a model&#8217;s performance in a language is capped by the data available, not by the architecture. Non-English languages hit it first. English is used on 49.5% of websites with a known content language (W3Techs, October 2026), and the web is the raw material most models learn from.<\/p>\n<p>This post breaks multilingual training data problems into four types, shows how to diagnose yours, and maps each one to a fix. A quick note on transparency: Ulatus, which provides data for AI services, publishes this article. So we flag where our own offering appears and label our judgements as such. If you build or buy data for AI, start your budget with the diagnosis.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"What_Is_the_Multilingual_Training_Data_Bottleneck\"><\/span>What Is the Multilingual Training Data Bottleneck?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>The multilingual training data bottleneck is the gap between the language data a model needs and the data that actually exists, is clean and reflects local culture. It limits how well a large language model reads, reasons and stays safe outside English.<\/p>\n<p>A <a href=\"https:\/\/en.wikipedia.org\/wiki\/Large_language_model\">large language model<\/a> learns from text in stages. Pretraining absorbs raw text, instruction tuning teaches it to follow requests, and reinforcement learning from human feedback (RLHF) steers it toward answers people prefer. A low-resource language has little digital text or few labeled datasets, which describes most of the more than 7,100 living languages in Ethnologue&#8217;s 2025 edition.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"The_four_forms_it_takes\"><\/span>The four forms it takes<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Each form has a different cause, so each one needs a different fix.<\/p>\n<ul>\n<li><strong>Volume:<\/strong> not enough text, speech or labeled examples in the language.<\/li>\n<li><strong>Quality:<\/strong> the data exists but is noisy, mislabeled, duplicated or machine-generated.<\/li>\n<li><strong>Cultural alignment:<\/strong> the data is fluent but reflects translated English assumptions, not local knowledge, norms or safety expectations.<\/li>\n<li><strong>Evaluation:<\/strong> you cannot measure progress because test sets are translated, thin or missing.<\/li>\n<\/ul>\n<h2><span class=\"ez-toc-section\" id=\"Why_Is_Multilingual_Training_Data_the_Bottleneck_for_LLMs_in_2026\"><\/span>Why Is Multilingual Training Data the Bottleneck for LLMs in 2026?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Multilingual training data is the bottleneck because web text skews heavily toward English, tokenizers charge non-English speakers more, and alignment data (instructions, preferences, safety examples) barely exists outside a few languages. Compute alone does not close gaps that targeted data can.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"Skewed_web_data_and_the_long_tail\"><\/span>Skewed web data and the long tail<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>The web&#8217;s English skew flows straight into crawled training corpora, which leaves the long tail of smaller languages with the least text. The raw material is getting better. The authors of FineWeb2, a 2025 Hugging Face corpus covering more than 1,000 languages, report that language-specific filtering and deduplication (removing repeated text) produce better non-English models than earlier corpora did. But covering a language is not the same as having enough of it.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"The_tokenization_tax\"><\/span>The tokenization tax<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Many languages pay a token tax: the same meaning takes more tokens, so it costs more and fits less into a model&#8217;s context window. A tokenizer splits text into units called tokens using schemes such as <a href=\"https:\/\/en.wikipedia.org\/wiki\/Byte-pair_encoding\">byte-pair encoding<\/a>, which were tuned mostly on English. A 2025 study of more than 200 languages found that non-Latin and morphologically complex languages often cost 3 to 5 times as many tokens as English. For one concrete case, LILT, a language-data vendor, found in a 2026 analysis that Arabic needed nearly 3 times as many tokens per dialogue as English (4,558 against 1,572).<\/p>\n<h3><span class=\"ez-toc-section\" id=\"Alignment_data_scarcity\"><\/span>Alignment data scarcity<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Instruction, preference and safety data are scarce because humans have to write them. Cohere&#8217;s Aya dataset, released in February 2024, needed an open-science community to produce 204,000 human-written prompt and completion pairs across 65 languages. Global-MMLU, published at ACL 2025, still relied on professional and community annotators to cover 42 languages. And a 2026 systematic review of safety alignment in low-resource languages found that safety weakens consistently as a language&#8217;s resource level drops.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"Why_compute_alone_does_not_fix_it\"><\/span>Why compute alone does not fix it<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Here&#8217;s the honest counterargument. LILT&#8217;s analysis, a vendor study on one multi-turn benchmark in Arabic, German and Korean, attributed 72.1% to 87.3% of non-English errors to model limitations and only 10.6% to 25.6% to data artifacts. So most errors sit in the model itself. Yet LILT&#8217;s own remedy is targeted post-training (further training of an existing model) on high-quality multilingual data. Few teams can redesign a foundation model, but every team controls what it trains on next. Data is the lever you can actually pull.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"How_Do_You_Diagnose_Which_Data_Bottleneck_You_Actually_Have\"><\/span>How Do You Diagnose Which Data Bottleneck You Actually Have?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Diagnose by matching symptoms to tests. Run a per-language evaluation slice, a language-ID audit, a token-cost check and a cultural red-team. Each exposes a different bottleneck: volume, quality, cultural alignment or evaluation. Diagnosing first stops you buying more of the wrong data.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"Quick_tests_worth_running_this_week\"><\/span>Quick tests worth running this week<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>A language-ID audit samples a few hundred documents per language and checks that they really are in that language. The evidence for doing this is older, but it still teaches something: a 2022 TACL audit of 205 web-crawled corpora found at least 15 were entirely erroneous, and 82 were mislabeled or used nonstandard language codes.<\/p>\n<p>A token-cost check compares tokens per equivalent request across languages. A per-language eval slice splits scores by language and task, because averages hide failures. On one editing task, for example, LILT saw Arabic and Korean collapse to 32% to 37% while German actually beat English.<\/p>\n<p>Picture a team whose model scores about 70% in English but about 40% in a priority language, the kind of gap MMLU-ProX found. Before buying anything, it runs the per-language slice and the language-ID audit. Weakness on every task points to volume, mislabeled text points to quality, and good scores alongside user complaints point to evaluation.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"Why_diagnosing_before_buying_saves_budget\"><\/span>Why diagnosing before buying saves budget<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Throwing volume at a cultural alignment problem wastes money, because more translated text will not teach a model local norms. A day of diagnostics tells you whether you need more data, cleaner data, better-aligned data or better tests. Those are four very different purchase orders.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"How_Can_LLM_Builders_Overcome_Multilingual_Data_Shortages\"><\/span>How Can LLM Builders Overcome Multilingual Data Shortages?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Match the remedy to the bottleneck. Collect targeted, consented data for volume gaps, filter and review for quality, commission native-written instruction and preference data for alignment, and build culturally grounded test sets for evaluation. Across all four, put humans between synthetic data and your model, and pilot on two languages before scaling.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"1_Fix_volume_targeted_collection_community_corpora_and_consent\"><\/span>1. Fix volume: targeted collection, community corpora and consent<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Start where your gap is largest and your users are most valuable. Check public resources first: if you serve India, the Bhashini platform hosts datasets and models you may be able to use before commissioning anything new. Then combine open corpora such as FineWeb2 with licensed material and custom <a href=\"https:\/\/www.ulatus.com\/ai-training-data\/data-collection-services\">multilingual data collection<\/a>, where contributors are vetted, demographics are balanced and consent is documented. Record provenance from day one. Unclear provenance and licensing create compliance risk, and EU rules now require general-purpose AI providers to summarise their training content.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"2_Fix_quality_language-ID_filtering_deduplication_and_native_review\"><\/span>2. Fix quality: language-ID filtering, deduplication and native review<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Filter by language identification, remove near-duplicates, and adapt the pipeline to each language instead of applying one global setting (the point FineWeb2&#8217;s authors make). Then add a human layer, because automated filters struggle most where data is scarcest. A native speaker reviewing a sample will catch what a classifier won&#8217;t.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"3_Fix_alignment_native-written_not_just_translated\"><\/span>3. Fix alignment: native-written, not just translated<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Why not just translate English instruction sets? It&#8217;s cheap, sure, but it imports English assumptions about names, laws, food, politeness and risk. Global-MMLU showed that model rankings shift on culturally sensitive questions and that machine-translated English tests carry Western-centric bias. The same logic plausibly applies to training data. So commission instruction, preference and safety examples written by native speakers, using professional <a href=\"https:\/\/www.ulatus.com\/ai-training-data\/data-annotation-services\">data annotation<\/a> with multi-pass review and agreement checks (how often independent reviewers give the same label). Safety examples need particular care. In a 2026 study of four African languages (Afrikaans, Kiswahili, isiXhosa and isiZulu), human red-teamers raised average jailbreak rates from 59.8% to 75.8% compared with automated translation.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"4_Fix_evaluation_per-language_culturally_grounded_test_sets\"><\/span>4. Fix evaluation: per-language, culturally grounded test sets<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Build test sets natively, not by translating English benchmarks. Include safety probes and long, multi-turn conversations, since LILT found non-English accuracy dropped sharply at 6 to 10 turns while English held steady. Use <a href=\"https:\/\/www.ulatus.com\/ai-training-data\/data-evaluation-services\">native-speaker model evaluation<\/a> to judge factuality, bias and cultural appropriateness, and keep these sets out of your training data.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"Guardrail_for_all_four_translate_then_post-edit\"><\/span>Guardrail for all four: translate, then post-edit<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Synthetic and machine-translated data touch every bottleneck. They&#8217;re useful, but they aren&#8217;t free. A 2024 EMNLP study on Indian languages found that models pretrained on filtered translationese (text that reads as translated) showed relative drops of only 0.87% on understanding tasks and 2.35% on generation tasks versus clean data, and the gap narrowed with a little clean text. Most of those models were small, so treat the result as encouraging rather than settled.<\/p>\n<p>For communities with little authentic text, the risk is sharper. A 2026 position paper argues that model collapse, where models degrade by training on generated output, threatens low-resource communities most. So keep a fixed floor of native data, tag synthetic data so you can test with and without it, and put <a href=\"https:\/\/www.ulatus.com\/machine-translation\/mtpe\">human post-editing of machine translation<\/a> between the translation engine and your training set.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"Build_Crowdsource_Synthesize_or_Partner_Which_Sourcing_Model_Fits\"><\/span>Build, Crowdsource, Synthesize or Partner: Which Sourcing Model Fits?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Build in-house when you need tight control over a few languages, crowdsource for simple, high-volume tasks, synthesize to scale pretraining and early fine-tuning, and partner with a specialist when cultural fidelity, compliance or evaluation matter most. Most teams end up mixing models by training stage.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"Decision_rule_by_stage\"><\/span>Decision rule by stage<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<ul>\n<li><strong>Pretraining:<\/strong> synthetic and open corpora, with native review of samples.<\/li>\n<li><strong>Supervised fine-tuning:<\/strong> a native-written core, synthetic expansion afterwards.<\/li>\n<li><strong>RLHF and preference data:<\/strong> in-house or specialist partner; crowd only with strong QA.<\/li>\n<li><strong>Evaluation and safety:<\/strong> specialist or in-house, never synthetic alone.<\/li>\n<\/ul>\n<h3><span class=\"ez-toc-section\" id=\"When_a_specialist_partner_wins\"><\/span>When a specialist partner wins<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>A partner earns its place when you need many languages at once, rare languages you can&#8217;t hire for, or quality you can audit. Full disclosure: Ulatus, the publisher of this article, is one such partner. Our <a href=\"https:\/\/www.ulatus.com\/ai-training-data\">data for AI<\/a> services draw on a self-reported pool of more than 100,000 vetted annotators, linguists and specialists in over 125 countries, covering 200+ languages. Ask any vendor, us included, for sample outputs, agreement scores and a pilot before you commit.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"A_90-Day_Roadmap_to_Close_Your_Multilingual_Training_Data_Gap\"><\/span>A 90-Day Roadmap to Close Your Multilingual Training Data Gap<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Spend the first month measuring, the second piloting fixes on two languages, and the third scaling what worked while setting governance. Track three numbers throughout: per-language eval delta, token cost per request, and refusal or unsafe-output rate.<\/p>\n<h3><span class=\"ez-toc-section\" id=\"Days_1-30_audit_and_baseline\"><\/span>Days 1-30: audit and baseline<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<ol>\n<li>Rank your languages by revenue or user value and pick the top two.<\/li>\n<li>Run the four diagnostic tests from the table above (owner: ML lead).<\/li>\n<li>Record baselines: per-language scores, tokens per request, safety failures.<\/li>\n<\/ol>\n<h3><span class=\"ez-toc-section\" id=\"Days_31-60_pilot_fixes_on_two_priority_languages\"><\/span>Days 31-60: pilot fixes on two priority languages<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<ol start=\"4\">\n<li>Apply only the fix your diagnosis points to (owner: data lead).<\/li>\n<li>Commission a small native-written set, for example a few thousand instruction and preference examples, plus a native test set.<\/li>\n<li>Post-train on the new data and compare against your baseline. Targeted post-training, not retraining a foundation model, is what LILT&#8217;s 2026 analysis recommends.<\/li>\n<\/ol>\n<h3><span class=\"ez-toc-section\" id=\"Days_61-90_scale_evaluate_and_set_governance\"><\/span>Days 61-90: scale, evaluate and set governance<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<ol start=\"7\">\n<li>Extend what worked to the next language tier (owner: program manager).<\/li>\n<li>Put in place <a href=\"https:\/\/www.ulatus.com\/quality-management\">linguistic quality management<\/a>: reviewer selection, sampling plans and agreement thresholds.<\/li>\n<li>Document data provenance and consent for every dataset.<\/li>\n<\/ol>\n<p>Since 2 August 2026, the EU AI Office has been able to enforce the AI Act&#8217;s rules for general-purpose AI providers. Those rules have applied since 2 August 2025. Providers must publish a public summary of training content using the Commission&#8217;s July 2025 template, and models already on the market before then have until 2 August 2027, according to a Mayer Brown summary. Fines run up to EUR 15 million or 3% of global turnover, whichever is higher.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"Treat_Multilingual_Training_Data_as_a_Product_Not_a_Backlog_Item\"><\/span>Treat Multilingual Training Data as a Product, Not a Backlog Item<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Multilingual training data problems come in four types: volume, quality, cultural alignment and evaluation. Each needs a different fix, which is why diagnosing first beats buying first. A 24.3% benchmark gap and token bills often 3 to 5 times higher are reason enough to start now.<\/p>\n<p>So treat each language like a product, with an owner, a roadmap and metrics. Run the four-bottleneck diagnostic on your two highest-value non-English languages this week. Then ask Ulatus to scope a two-language pilot of native-speaker, linguist-reviewed data for AI, with sample outputs and agreement scores up front, so your next model release closes the gap instead of widening it.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>The Data Wall Hits Non-English First Ask the same 11,829 questions in 29 languages and Qwen2.5-72B scores 70.3% in English but only 40.1% in Swahili. That result comes from MMLU-ProX, a 2025 benchmark of 36 models that found gaps of up to 24.3% between high- and low-resource languages. And the pattern points to one place: [&hellip;]<\/p>\n","protected":false},"author":25,"featured_media":3237,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[464],"tags":[],"class_list":["post-3236","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-data-for-ai"],"acf":[],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.7 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It\" \/>\n<meta property=\"og:description\" content=\"The Data Wall Hits Non-English First Ask the same 11,829 questions in 29 languages and Qwen2.5-72B scores 70.3% in English but only 40.1% in Swahili. That result comes from MMLU-ProX, a 2025 benchmark of 36 models that found gaps of up to 24.3% between high- and low-resource languages. And the pattern points to one place: [&hellip;]\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/\" \/>\n<meta property=\"article:published_time\" content=\"2026-11-04T05:17:01+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.ulatus.com\/translation-blog\/wp-content\/uploads\/2026\/10\/banner-multilingual-training-data-bottleneck-llm-hourglass.png\" \/>\n\t<meta property=\"og:image:width\" content=\"1280\" \/>\n\t<meta property=\"og:image:height\" content=\"720\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Elizabeth Rodrigues\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@ulatus\" \/>\n<meta name=\"twitter:site\" content=\"@ulatus\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Elizabeth Rodrigues\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"10 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/\"},\"author\":{\"name\":\"Elizabeth Rodrigues\",\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/#\\\/schema\\\/person\\\/2c88b235c1a21525603db0395518c787\"},\"headline\":\"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It\",\"datePublished\":\"2026-11-04T05:17:01+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/\"},\"wordCount\":2088,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/banner-multilingual-training-data-bottleneck-llm-hourglass.png\",\"articleSection\":[\"Data for AI\"],\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/\",\"url\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/\",\"name\":\"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/banner-multilingual-training-data-bottleneck-llm-hourglass.png\",\"datePublished\":\"2026-11-04T05:17:01+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/#\\\/schema\\\/person\\\/2c88b235c1a21525603db0395518c787\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/#primaryimage\",\"url\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/banner-multilingual-training-data-bottleneck-llm-hourglass.png\",\"contentUrl\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/banner-multilingual-training-data-bottleneck-llm-hourglass.png\",\"width\":1280,\"height\":720,\"caption\":\"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/#website\",\"url\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/\",\"name\":\"\",\"description\":\"Translation Trends &amp; Insights\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/#\\\/schema\\\/person\\\/2c88b235c1a21525603db0395518c787\",\"name\":\"Elizabeth Rodrigues\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/8cbc606363a4bbd8eda6ccc8c6a750eae05a2de996ee4dc49e0c776598bb1d45?s=96&d=mm&r=g\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/8cbc606363a4bbd8eda6ccc8c6a750eae05a2de996ee4dc49e0c776598bb1d45?s=96&d=mm&r=g\",\"contentUrl\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/8cbc606363a4bbd8eda6ccc8c6a750eae05a2de996ee4dc49e0c776598bb1d45?s=96&d=mm&r=g\",\"caption\":\"Elizabeth Rodrigues\"},\"url\":\"https:\\\/\\\/www.ulatus.com\\\/translation-blog\\\/author\\\/elizabeth-rodrigues\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/","og_locale":"en_US","og_type":"article","og_title":"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It","og_description":"The Data Wall Hits Non-English First Ask the same 11,829 questions in 29 languages and Qwen2.5-72B scores 70.3% in English but only 40.1% in Swahili. That result comes from MMLU-ProX, a 2025 benchmark of 36 models that found gaps of up to 24.3% between high- and low-resource languages. And the pattern points to one place: [&hellip;]","og_url":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/","article_published_time":"2026-11-04T05:17:01+00:00","og_image":[{"width":1280,"height":720,"url":"https:\/\/www.ulatus.com\/translation-blog\/wp-content\/uploads\/2026\/10\/banner-multilingual-training-data-bottleneck-llm-hourglass.png","type":"image\/png"}],"author":"Elizabeth Rodrigues","twitter_card":"summary_large_image","twitter_creator":"@ulatus","twitter_site":"@ulatus","twitter_misc":{"Written by":"Elizabeth Rodrigues","Est. reading time":"10 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#article","isPartOf":{"@id":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/"},"author":{"name":"Elizabeth Rodrigues","@id":"https:\/\/www.ulatus.com\/translation-blog\/#\/schema\/person\/2c88b235c1a21525603db0395518c787"},"headline":"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It","datePublished":"2026-11-04T05:17:01+00:00","mainEntityOfPage":{"@id":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/"},"wordCount":2088,"commentCount":0,"image":{"@id":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#primaryimage"},"thumbnailUrl":"https:\/\/www.ulatus.com\/translation-blog\/wp-content\/uploads\/2026\/10\/banner-multilingual-training-data-bottleneck-llm-hourglass.png","articleSection":["Data for AI"],"inLanguage":"en-US","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#respond"]}]},{"@type":"WebPage","@id":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/","url":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/","name":"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It","isPartOf":{"@id":"https:\/\/www.ulatus.com\/translation-blog\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#primaryimage"},"image":{"@id":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#primaryimage"},"thumbnailUrl":"https:\/\/www.ulatus.com\/translation-blog\/wp-content\/uploads\/2026\/10\/banner-multilingual-training-data-bottleneck-llm-hourglass.png","datePublished":"2026-11-04T05:17:01+00:00","author":{"@id":"https:\/\/www.ulatus.com\/translation-blog\/#\/schema\/person\/2c88b235c1a21525603db0395518c787"},"breadcrumb":{"@id":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#primaryimage","url":"https:\/\/www.ulatus.com\/translation-blog\/wp-content\/uploads\/2026\/10\/banner-multilingual-training-data-bottleneck-llm-hourglass.png","contentUrl":"https:\/\/www.ulatus.com\/translation-blog\/wp-content\/uploads\/2026\/10\/banner-multilingual-training-data-bottleneck-llm-hourglass.png","width":1280,"height":720,"caption":"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It"},{"@type":"BreadcrumbList","@id":"https:\/\/www.ulatus.com\/translation-blog\/why-multilingual-training-data-is-the-new-bottleneck-for-llm-builders-and-how-to-overcome-it\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.ulatus.com\/translation-blog\/"},{"@type":"ListItem","position":2,"name":"Why Multilingual Training Data Is the New Bottleneck for LLM Builders. And How to Overcome It"}]},{"@type":"WebSite","@id":"https:\/\/www.ulatus.com\/translation-blog\/#website","url":"https:\/\/www.ulatus.com\/translation-blog\/","name":"","description":"Translation Trends &amp; Insights","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.ulatus.com\/translation-blog\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Person","@id":"https:\/\/www.ulatus.com\/translation-blog\/#\/schema\/person\/2c88b235c1a21525603db0395518c787","name":"Elizabeth Rodrigues","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/secure.gravatar.com\/avatar\/8cbc606363a4bbd8eda6ccc8c6a750eae05a2de996ee4dc49e0c776598bb1d45?s=96&d=mm&r=g","url":"https:\/\/secure.gravatar.com\/avatar\/8cbc606363a4bbd8eda6ccc8c6a750eae05a2de996ee4dc49e0c776598bb1d45?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/8cbc606363a4bbd8eda6ccc8c6a750eae05a2de996ee4dc49e0c776598bb1d45?s=96&d=mm&r=g","caption":"Elizabeth Rodrigues"},"url":"https:\/\/www.ulatus.com\/translation-blog\/author\/elizabeth-rodrigues\/"}]}},"_links":{"self":[{"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/posts\/3236","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/users\/25"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/comments?post=3236"}],"version-history":[{"count":1,"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/posts\/3236\/revisions"}],"predecessor-version":[{"id":3238,"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/posts\/3236\/revisions\/3238"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/media\/3237"}],"wp:attachment":[{"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/media?parent=3236"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/categories?post=3236"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ulatus.com\/translation-blog\/wp-json\/wp\/v2\/tags?post=3236"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}