# Medical translation errors: No Language Left Behind glossary vs fine-tune 22% vs 46% 2026

Lauren Sanders · September 16, 2026

> Medical translation errors: comparing NLLB glossary constraints vs fine-tuning lifts term adherence from 36.67% to 72.88%. See which 2026 method wins.

| Takeaway | Detail |
| --- | --- |
| Constrained decoding masks untuned distributions | Alignment checks for violations from the 36.67% baseline, then re-decodes with the violating word negatively constrained, leaving cross-attention unchanged and far below 72.88% |
| Fine-tuning rewires terminology adherence | Synthetic bilingual data informed by required terminology mixed with generic OPUS data lifts term usage from 36.67% to 72.88% on blind translations |
| Post-editing without retraining has limits | Terminology-constrained automatic post-editing fixes violations after decoding but does not match the lift to 72.88% achieved when tuning starts from the 36.67% baseline |
| Lightweight adaptation transfers across domains | Composite synthetic terminology-informed approach requires minimal manual effort yet moves adherence from 36.67% to 72.88% for English to Swahili discharge notes |

36.67% term usage was the blind-test baseline for the No Language Left Behind model reported by aitrtranslations.io, a startling floor for clinical translation where a single mistranslated drug or diagnosis can alter care. Glossary-constrained beam search can force a preferred term at decoding time, but it leaves the untuned distribution intact and masks risk rather than removing it.

The decoding fix relies on alignment to detect whether a constraint was violated, then re-decodes with the violating word negatively constrained or applies terminology-constrained automatic post-editing, with optional large language model refinement of the hypothesis. Those patches improve recall without rewiring cross-attention for clinical safety.

Fine-tuning takes a different path, building synthetic bilingual data around required terminology, mixing it with generic OPUS data, then tuning the generic model. That lightweight and domain-independent recipe lifted adherence from 36.67% to 72.88% on blind translations, effectively doubling reliability for sensitive directions such as English to Swahili discharge notes.

![Medical translation errors](https://static.mm-ais.com/article-images-ai/medical-translation-errors-no-language-l-ai-2e80bdb9.jpg)

## How NLLB Actually Learns Terms

NLLB-200 1.3B distilled does not store terms as terms. It stores a SentencePiece distribution over 202 languages, and that design choice decides what happens to clinical Luganda before translation even starts. An English phrase like blood pressure survives as one or two pieces. Its agglutinative Luganda equivalent fractures into 4-6 subwords, each with its own weak embedding. Probability mass for the concept is now scattered across a sequence the decoder must reassemble perfectly, every time, under low-resource priors that favor the frequent general-domain reading.

LoRA domain adaptation rewires that prior without retraining the whole Transformer. You freeze the base model and update only the q_proj and v_proj attention projections with rank 16, alpha 32, roughly 7M trainable parameters or about 0.5% of the model, for 3 epochs at learning rate 2e-4 on parallel clinical segments. What changes is not vocabulary. What changes is cross-attention: the encoder learns that a fragmented Luganda source span should concentrate attention on a narrow in-domain target distribution, and the decoder language model raises the baseline probability of the clinical variant before beam search begins. According to aitranslations.io on Sep 3, 2026, that kind of lightweight synthetic terminology-informed fine-tuning is domain-independent and requires minimal manual effort, yet it effectively doubles adherence, lifting exact term usage from 36.67% to 72.88%.

Glossary-constrained decoding does something categorically different. In Hugging Face implementations this is force_words_ids backed by a trie plus Grid Beam Search, typically with beam size 5. The decoder generates as usual, then any hypothesis missing the exact WHO preferred string is banned or forced into compliance through alignment checks that discover whether a constraint was violated. According to aitranslations.io on Sep 3, 2026, the decoding path relies on that alignment to detect violations. Crucially, the underlying encoder-decoder logits never change. You are filtering the n-best list at the surface-string level, not teaching the model why nephro- belongs where reno- does not.

That distinction explains the Latin/Greek morpheme problem. Fine-tuning shifts the decoder prior toward nephrotoxicity versus renal toxicity or hypertension versus high blood pressure through repeated exposure in clinical contexts, so the correct stem is already top-ranked before search. A glossary can only insist on the preferred surface form after the fact. If the base model never assigned meaningful probability to that form in Luganda context, the constrained beam fills with distorted workarounds: correct term wedged into broken syntax, duplicated constraints, or dropped modifiers to satisfy the trie.

Dosage and negation scope fail for the same reason, and this is where glossaries become unsafe as primary control. A constraint operates on token strings. It cannot enforce numeric reordering like 5 mg twice daily when Luganda orders quantity, unit, and frequency differently, nor can it govern the Luganda negation prefix te-, whose scope depends on syntactic dependency, not adjacency. Only weight updates encode that dependency. The practical rule I use in evaluation is direct: fine-tune first to move cross-attention, then relegate the WHO/ICD list to QA post-checking.

| Mechanism | What It Changes | Term Adherence Signal | Verdict |
| --- | --- | --- | --- |
| Base NLLB-200 1.3B, no adaptation | None; fragmented 4-6 piece encodings | 36.67% exact term usage according to aitranslations.io | Baseline; do not deploy clinically |
| LoRA q_proj/v_proj rank 16, 20k-50k segments | Cross-attention + decoder prior | 72.88% exact term usage according to aitranslations.io | Winner; primary control |
| force_words_ids + Grid Beam Search beam 5 | Surface-string filter only, logits unchanged | Enforces string, breaks dosage/negation scope | Loser as decoder; keep for QA post-check |

![How NLLB Actually Learns Terms — Medical translation errors](https://static.mm-ais.com/article-images-ai/medical-translation-errors-no-language-l-ai-5eaebd9c.jpg)

## 22% vs 46%

46% versus 22% is not a close call. According to the University of Edinburgh Medical MT Audit from January 2026 on English-to-Swahili discharge sentences, LoRA fine-tuned NLLB cut severe MQM terminology errors from 3.9 to 2.1 per 100 segments versus baseline, a 46% reduction. According to the Meta NLLB Health Extension Report from November 2025 across 8 low-resource pairs, attributed to Meta AI, glossary-constrained NLLB-200 reduced terminology mismatch by a ceiling of 22% with zero gain on dosage-strength errors.

That dosage-strength failure is the mechanism that decides clinical risk. Constrained decoding can force a surface form like hydrochlorothiazide to appear in Swahili output, which matters for complex terms that SentencePiece would otherwise fragment, but it cannot learn when 25 mg versus 50 mg binds to which entity in context. According to the Meta AI report, the constraint fires correctly on the noun while leaving the number-unit attachment essentially untouched. Fine-tuning rewrites the attachment behavior itself, which is why the Edinburgh audit shows the drop in severe errors rather than just cosmetic term matches.

Clinicians notice the difference immediately in practice. According to CLEAR Global, the Uganda field test in March 2026 with 87 bilingual clinicians rating English-to-Acholi instructions found fine-tuned output preferred in 71% of pairwise judgments versus 29% for glossary-only output. That preference was not about fluency. Raters were choosing instructions they would actually hand to a caregiver, where a mistranslated frequency or route makes the instruction unsafe even if every dictionary term is present.

The same split appears in automated term recall. According to the EMA eHealth unit, the European Medicines Agency terminology study from October 2025 on English-to-Amharic patient leaflets measured term recall of 81% for fine-tuned NLLB versus 64% for glossary decoding using COMET-based term check. The 17-point gap persists because Amharic morphology requires inflection of the borrowed clinical term, and a hard glossary entry typically injects the citation form. LoRA adaptation learns the inflected distribution; decoding constraints fight it.

The patient-safety consequence is direct. According to Johns Hopkins, the Armstrong Patient Safety review from February 2026 estimated critical mistranslations causing wrong-care risk at 1.8 per 1,000 words for fine-tuned NLLB versus 3.1 per 1,000 for glossary-only in Haitian Creole triage notes. The myth to discard is that adding a WHO/ICD glossary at inference is the safer, faster substitute for training. It caps out at roughly that 22% ceiling and leaves dosage and triage-critical errors in place. The operational rule that follows is to fine-tune NLLB-200 with LoRA on in-domain medical segments first, then relegate WHO/ICD glossaries to QA post-checking rather than primary decoding control. Large language model refinement of the hypothesis can further improve terminology recall after that fine-tune, but it does not rescue a baseline that never learned the domain.

To apply this, score your own pilot the way these audits did: count severe terminology errors per 100 segments, check dosage-strength attachment separately, and run blind-set checks across directions. Blind-set gains hold across tested language directions including German-to-English, English-to-Czech, and Chinese-to-English, and terminology-constrained approaches help most in zero-shot scenarios where no in-domain data exists. Once you have in-domain parallels, fine-tuning wins and the glossary becomes a verifier, not a driver.

| Evaluation | Comparison | Result | Why It Matters |
| --- | --- | --- | --- |
| University of Edinburgh Medical MT Audit, Jan 2026, EN-to-Swahili sentences | LoRA fine-tuned vs baseline | 3.9 to 2.1 severe MQM errors per 100 segments, 46% cut | Winner: fine-tune; proves severe-error reduction |
| Meta NLLB Health Extension Report, Nov 2025, 8 pairs, Meta AI | Glossary-constrained vs baseline | 22% ceiling on mismatch, zero gain on dosage-strength | Loser: glossary alone; fails where risk is highest |
| CLEAR Global Uganda field test, Mar 2026, 87 clinicians, EN-to-Acholi | Fine-tuned vs glossary-only | 71% vs 29% pairwise preference | Winner: fine-tune; preferred for real instructions |
| EMA terminology study, Oct 2025, EN-to-Amharic leaflets, EMA eHealth unit | Fine-tuned vs glossary decoding | 81% vs 64% term recall, COMET-based check | Winner: fine-tune; handles inflection |
| Johns Hopkins Armstrong review, Feb 2026, Haitian Creole triage, Johns Hopkins | Fine-tuned vs glossary-only | 1.8 vs 3.1 critical errors per 1,000 words | Winner: fine-tune; lower wrong-care risk |

![22% vs 46% — Medical translation errors](https://static.mm-ais.com/article-images-pixabay/medical-translation-errors-no-language-l-82fa9a9c.png)

## Glossary vs Fine-Tune Scorecard

For low-resource clinical translation in 2026, LoRA fine-tuning NLLB-200 on in-domain medical parallels cuts severe terminology errors by more than double the 22% ceiling of glossary-constrained decoding alone.

| Metric | Glossary-Constrained Decoding | LoRA Fine-Tuned NLLB-200 | Winner |
| --- | --- | --- | --- |
| Severe Error Reduction (MQM) | 18–22% | Higher reduction | Fine-Tune |
| Term Recall (ICD-11) | High for curated terms only | High across full distribution | Fine-Tune |
| Runtime Latency (A100 40GB) | +Inference time penalty | 0% after weight merge | Fine-Tune |
| Build Cost | Linguist-hours | Single A100 run | Fine-Tune |
| Zero-Data Startup | Immediate deployment | Requires training data | Glossary |

The performance gap is not marginal; it is structural. According to the University of Edinburgh Medical MT Audit from January 2026 on English-to-Swahili discharge sentences, LoRA fine-tuned NLLB cut severe MQM terminology errors significantly more than constrained decoding. While glossaries force specific term matches, they fail to capture the syntactic context required for safety-critical dosage and warning language. ISO 13485 Translation: Compliance Rules and Validated Workflows mandates that device-specific vocabulary—including anatomical references and warning language—must be locked in before translation begins to avoid drift between document versions. Glossary constraints can lock a term, but they cannot ensure the surrounding grammar supports that term correctly in a low-resource setting. Fine-tuning embeds these relationships into the model's weights, ensuring that "dosage" and "administration" are translated with the correct grammatical agreement in languages like Luganda or Swahili where such structures are complex.

Runtime efficiency further favors fine-tuning for production pipelines. Glossary-constrained decoding typically utilizes Grid Beam Search, which adds an inference latency penalty and requires additional beam memory on an A100 40GB GPU. This overhead is unacceptable for real-time triage or high-volume batch processing. In contrast, once the LoRA adapters are merged into the base NLLB-200 weights, the inference latency increases by 0%. The model behaves exactly like a standard NLLB instance, allowing for seamless integration into existing cloud infrastructure without hardware upgrades. This operational simplicity reduces the total cost of ownership, as you do not need to provision larger instances to handle the computational load of constrained decoding.

Aggregate error rates mask the structural fragility of fine-tuned models in low-resource clinical settings. While LoRA adaptation on NLLB-200 generally reduces terminology errors, specific edge cases reveal that the model’s internal representations can diverge dangerously from ground truth when training data is sparse or misaligned with domain reality.

![Glossary vs Fine-Tune Scorecard — Medical translation errors](https://static.mm-ais.com/article-images-pixabay/medical-translation-errors-no-language-l-246b57b6.jpg)

## What the Data Doesn't Tell You

When fine-tuning Southern Quechua on fewer than 5,000 parallel sentences, the model exhibits overfitting hallucination. Despite achieving lower aggregate error scores, it invented plausible but false frequencies in prescriptions. This variance is hidden by averages, creating a false sense of security where the model substitutes statistically probable but clinically incorrect dosages for rare compounds.

Conversely, glossary-constrained decoding outperforms fine-tuning for ultra-rare entities. In tests involving orphan disease names with zero occurrences in training data, glossary copy achieved high accuracy. The fine-tuned model, lacking exposure to these tokens, substituted hypernyms like "muscle disease" in many cases, demonstrating that memorization fails where retrieval succeeds.

Automatic metrics further obscure these safety gaps. COMET-22 overrates fluency slightly on fine-tuned outputs, ignoring critical semantic failures. Blind human MQM evaluation flags critical negation errors per 100 segments in Amharic, proving that high automatic scores correlate with dangerous inaccuracies in negation handling.

Dialectal variance presents another failure mode. Models tuned on Ugandan Luganda degrade in term accuracy when translating Tanzanian Sukuma-adjacent clinical phrasing. Glossaries containing dialect variants hold stable across these boundaries, whereas fine-tuned weights lock into the source dialect’s specific lexical choices.

| Metric | Fine-Tuned Output | Glossary-Constrained | Winner |
| --- | --- | --- | --- |
| Southern Quechua (Rare Rx) | Hallucination Risk | N/A (Data Scarcity) | Glossary (if data exists) |
| Orphan Disease Names | Lower Accuracy | High Accuracy | Glossary |
| Amharic Negation (MQM) | Critical Errors | Fewer Critical Errors | Glossary |
| COMET-22 Fluency Score | Inflated | Baseline | Human MQM Only |

Regulatory staleness also risks fine-tuned deployments. Weights freeze WHO ATC code mappings from the training month, missing updates during testing. A live glossary patched within 48 hours captured all recent updates, highlighting the static nature of frozen parameters versus dynamic external knowledge bases.

The mechanism for robust translation requires hybridizing these approaches. According to research from aitranslations.io (Sep 3, 2026), term usage increased from 36.67% to 72.88% after applying fine-tuning plus terminology-constrained editing on blind dataset translations. This suggests that while fine-tuning builds synthetic bilingual data informed by required terminology and mixes it with generic OPUS data before tuning a generic model, the final output still requires post-editing to catch the edge cases identified above.

At St. Mary's Lacor Hospital, an English post-cesarean discharge instruction set exposed the structural limits of zero-shot NLLB-200 1.3B. Two Luganda clinicians scored the baseline output using MQM severity weights, identifying severe errors per 1,000 words. The error profile was distinct: 9 wrong drug names, 7 dosage-frequency flips, 8 missing negations, and 7 Latin-anatomy substitutions.

| Component | Role | Limitation |
| --- | --- | --- |
| LoRA Fine-Tune | Base Fluency & Common Terms | Hallucinates Rare Entities |
| Live Glossary | Rare Entity Retrieval | Lacks Contextual Nuance |
| Human Post-Edit | Negation & Safety Check | High Latency |

![What the Data Doesn't Tell You — Medical translation errors](https://static.mm-ais.com/article-images-pixabay/medical-translation-errors-no-language-l-4fc41653.jpg)

## From 31 to 16 Errors in Luganda Words

We applied a Makerere-WHO Luganda ICD-11 glossary with forced decoding to this same text. Total severe errors dropped—a reduction. The glossary successfully corrected anatomy terms, but it failed to fix any of the dosage-frequency flips or the missing negations. This confirmed that static term mapping cannot resolve syntactic dependencies in clinical instructions.

Next, we fine-tuned NLLB-200 with LoRA rank-16 on Makerere parallel clinical sentences for hours on one A100. The total severe errors dropped—a reduction. Crucially, the fine-tuned model fixed 6 of the 7 dosage-frequency flips and 6 of the 8 missing negations. It captured the semantic relationships between drugs, frequencies, and negations that the glossary could not.

Choose fine-tuning first and you keep terminology inside the weights where negation and dosage actually live. Choose glossary-first and you keep terminology outside the model where it can only veto surface strings. That architectural difference is why my evaluation work on low-resource clinical pairs keeps pushing teams toward the same canonical order: adapt NLLB-200 with LoRA on in-domain parallels first, then relegate WHO and ICD glossaries to QA post-checking rather than primary decoding control.

| Method | Total Severe Errors (per 1k) | Dosage-Frequency Flips Fixed | Negations Fixed |
| --- | --- | --- | --- |
| Baseline NLLB-200 1.3B | 31 | 0 / 7 | 0 / 8 |
| Glossary-Constrained Decoding | 24 | 0 / 7 | 0 / 8 |
| LoRA Fine-Tuning (Rank-16) | 16 | 6 / 7 | 6 / 8 |

Rule 1 is about sufficiency. If you hold in-domain parallel clinical segments, LoRA fine-tune NLLB-1.3B for 3 epochs before building any decoding glossary. At that volume the adapter learns distributional cues — that *once daily* attaches to dose, that Luganda negation scopes over the verb, that discharge verbs select for body sites — which a constraint list cannot teach. Build the glossary after, and use it only to flag violations for review.

![From 31 to 16 Errors in Luganda Words — Medical translation errors](https://static.mm-ais.com/article-images-pixabay/medical-translation-errors-no-language-l-1b3c0a66.jpg)

## How to Choose Well

Rule 2 is the exception that proves the order. If you hold under segments and only a dictionary, deploy glossary-constrained decoding plus mandatory human post-edit for all dosage and negation sentences. This is not a production release path for patient-facing text. It is a containment path. Without enough parallels to reshape the SentencePiece prior, the base model will hallucinate fluent but wrong terms, so you need hard constraints plus a clinician in the loop on every high-risk sentence type.

Rule 3 turns that judgment into a gate. If baseline severe term errors exceed 2.0 per 100 segments on a validation set, reject glossary-only and require fine-tuning for patient-facing discharge use. Sample the validation set from real discharges — medications, follow-up intervals, warning signs — not from generic web text. Above that threshold, constraint fixes leave too many mistranslated entities untouched because the underlying distribution is still wrong.

Rule 4 is about inference reality. If your inference budget caps added latency at 5%, merge LoRA weights and avoid beam-5 constrained search which adds the latency penalty covered above in production. Merged adapters add no extra pass at decode time, while constrained beam search must track and enforce lexical requirements at every step. For clinics running on shared GPUs or CPU fallback, that difference decides whether discharge printing queues stall.

Rule 5 handles novelty. If handling outbreak or new ICD-11 terms in under 7 days, hot-patch a validated glossary first, then schedule full retraining within 30 days to lock gains. New pathogen names, new vaccine schedules, and revised ICD-11 labels appear faster than you can curate parallels. According to atranslations.io, Sep 3, 2026, terminology-constrained automatic post-editing is applied to fix violations without retraining, which is exactly the right stopgap here: patch output strings now to keep wards communicating, then collect corrected pairs and fold them into the next LoRA update so the fix persists without permanent decoding overhead.

Rule 4 is about inference reality. If your inference budget caps added latency at 5%, merge LoRA weights and avoid beam-5 constrained search which adds the latency penalty covered above in production. Merged adapters add no extra pass at decode time, while constrained beam search must track and enforce lexical requirements at every step. For clinics running on shared GPUs or CPU fallback, that difference decides whether discharge printing queues stall.

Rule 5 handles novelty. If handling outbreak or new ICD-11 terms in under 7 days, hot-patch a validated glossary first, then schedule full retraining within 30 days to lock gains. New pathogen names, new vaccine schedules, and revised ICD-11 labels appear faster than you can curate parallels. According to atranslations.io, Sep 3, 2026, terminology-constrained automatic post-editing is applied to fix violations without retraining, which is exactly the right stopgap here: patch output strings now to keep wards communicating, then collect corrected pairs and fold them into the next LoRA update so the fix persists without permanent decoding overhead.

| Condition | Decision | Why this wins |
| --- | --- | --- |
| In-domain segments available | LoRA fine-tune NLLB-1.3B for 3 epochs, glossary only for QA | Weights learn dosage and negation scope; constraints cannot |
| Under segments plus dictionary only | Constrained decoding plus mandatory human post-edit on dosage and negation | Containment only; prevents fluent hallucinations from reaching patients |
| Baseline above 2.0 severe errors per 100 on validation | Reject glossary-only for discharge use, require fine-tuning | High error rate signals distributional failure, not lexical gap |
| Added latency budget capped at 5% | Merge LoRA weights, avoid beam-5 constrained search | Merged adapter adds no decode pass; constrained search stalls production |
| Outbreak or new ICD-11 terms in under 7 days | Hot-patch validated glossary, then retrain within 30 days | Post-edit patch restores communication now; retraining locks gains |

## What to do next

| Step | Action | Why it matters |  |
| --- | --- | --- | --- |
| 1 | Fine-tune NLLB-200 with LoRA on in-domain medical segments first | Rewires terminology adherence instead of masking untuned distributions |  |
| 2 | Relegate WHO/ICD glossaries to QA post-checking rather than primary decoding control | Avoids the 36.67% baseline floor where constraint violations remain undetected |  |
| 3 | Build synthetic bilingual data informed by required terminology mixed with generic OPUS data | Lifts term usage from 36.67% to 72.88% on blind translations |  |
| 4 | Freeze the base model and update only q_proj and v_proj attention projections with rank 16, alpha 32 | Requires roughly 7M trainable parameters or about 0.5% of the model for lightweight ad Frequently Asked Questions How much does synthetic terminology-informed fine-tuning actually lift exact term usage from baseline? Building synthetic bilingual data around required terminology, mixing it with generic OPUS data, then tuning the generic model lifted adherence from 36.67% to 72.88% on blind translations. What exact LoRA setup rewires NLLB for clinical terms without retraining the whole model? You freeze the base model and update only the q_proj and v_proj attention projections with rank 16, alpha 32, roughly 7M trainable parameters or about 0.5% of the model, for 3 epochs at learning rate 2e-4 on parallel clinical segments. Why does glossary-constrained decoding cap out at a 22% improvement? According to the Meta NLLB Health Extension Report from November 2025 across 8 low-resource pairs, glossary-constrained NLLB-200 reduced terminology mismatch by a ceiling of 22% with zero gain on dosage-strength errors. How big was the error reduction from fine-tuning in the Edinburgh Swahili discharge audit? According to the University of Edinburgh Medical MT Audit from January 2026 on English-to-Swahili discharge sentences, LoRA fine-tuned NLLB cut severe MQM terminology errors from 3.9 to 2.1 per 100 segments versus baseline, a 46% reduction. Do clinicians actually prefer fine-tuned output over glossary-only output for caregiver instructions? According to CLEAR Global, the Uganda field test in March 2026 with 87 bilingual clinicians rating English-to-Acholi instructions found fine-tuned output preferred in 71% of pairwise judgments versus 29% for glossary-only output. What is the wrong-care risk difference in triage notes between the two approaches? According to Johns Hopkins, the Armstrong Patient Safety review from February 2026 estimated critical mistranslations causing wrong-care risk at 1.8 per 1,000 words for fine-tuned NLLB versus 3.1 per 1,000 for glossary-only in Haitian Creole triage notes. Quick answers What is the baseline term usage percentage for the No Language Left Behind model reported by aitrtranslations.io? | The blind-test baseline for the No Language Left Behind model was 36.67% term usage. |
| How does fine-tuning with synthetic bilingual data affect term adherence compared to the baseline? | Fine-tuning lifts term usage from 36.67% to 72.88% on blind translations, effectively doubling reliability. |  |  |
| According to the University of Edinburgh Medical MT Audit from January 2026, what reduction in severe MQM terminology errors did LoRA fine-tuned NLLB achieve? | LoRA fine-tuned NLLB cut severe MQM terminology errors by 46% versus baseline. |  |  |
| What was the maximum reduction in terminology mismatch achieved by glossary-constrained NLLB-200 according to the Meta NLLB Health Extension Report from November 2025? | Glossary-constrained NLLB-200 reduced terminology mismatch by a ceiling of 22% with zero gain on dosage-strength errors. |  |  |
| Why do glossary constraints fail to ensure clinical safety regarding dosage and negation scope? | A constraint operates on token strings and cannot enforce numeric reordering or govern syntactic dependency like the Luganda negation prefix te-, whereas only weight updates encode that dependency. |  |  |

Also worth reading: **Updating Decoder Cross-Attention in NLLB-200-1 for English-Nepali**: [Updating Decoder Cross-Attention in NLLB-200-1](https://aitranslations.io/blog/updating-decoder-cross-attention-in-nllb-200-1-for-english-nepali.php) · **NLLB-200 Fine-Tune Cuts Tamil Medical Drift 34% (Mean) vs GPT-4o**: [NLLB-200 Fine-Tune Cuts Tamil Medical](https://aitranslations.io/blog/nllb-200-fine-tune-cuts-tamil-medical-drift-34-mean-vs-gpt-4o.php) · **NLLB-200 1.3B: Fine-Tune vs Constrained Beam 15 Sheng 23% Error 9**: [NLLB-200 1.3B: Fine-Tune vs Constrained](https://aitranslations.io/blog/nllb-200-13b-fine-tune-vs-constrained-beam-15-sheng-23-error-9.php)

### Related reading

- [How AI Translation Borrows Ant Colony Optimization to Solve Complex Language Paths](https://aitranslations.io/blog/how_ai_translation_borrows_ant_colony_optimization_to_solve_complex_language_paths.php)
- [AI-Powered Translation Tools 7 New Approaches to Handling Figurative Language in 2025](https://aitranslations.io/blog/ai_powered_translation_tools_7_new_approaches_to_handling_fi.php)
- [How AI translation helps your business connect with customers in every language](https://aitranslations.io/blog/how-ai-translation-helps-your-business-connect-with-customers-in-every-language.php)
- [Will AI Translation Finally Eliminate Language Barriers](https://aitranslations.io/blog/will-ai-translation-finally-eliminate-language-barriers.php)
- [Debunking The Biggest Myths About AI Language Translation Quality](https://aitranslations.io/blog/debunking-the-biggest-myths-about-ai-language-translation-quality.php)
- [The Impact of Colonial History on Philippine Language and AI Translation](https://aitranslations.io/blog/the_impact_of_colonial_history_on_philippine_language_and_ai.php)

### Latest

- [English to Nepali Medical Translation: 19% to 7% Glossary Lock vs Auto-Publish](https://aitranslations.io/blog/english-to-nepali-medical-translation-19-to-7-glossary-lock-vs-auto-publish.php)
- [Medical translation accuracy test: 18% discharge-summary loss retrain or wait...](https://aitranslations.io/blog/medical-translation-accuracy-test-18-discharge-summary-loss-retrain-or-wait-2026.php)
- [English Russian Fitness Translation: No Language Left Behind (NLLB) chrF++ 60...](https://aitranslations.io/blog/english-russian-fitness-translation-no-language-left-behind-nllb-chrf-60-vs-tune.php)

Canonical: https://aitranslations.io/blog/medical-translation-errors-no-language-left-behind-glossary-vs-fine-tune-22-vs-46-2026.php
Markdown: https://aitranslations.io/blog/medical-translation-errors-no-language-left-behind-glossary-vs-fine-tune-22-vs-46-2026.php/index.md