2026 Quechua Spanish terminology: Adapter vs Full Fine-Tuning COMET/BLEU per 1,000

This guide compares adapter and full fine-tuning for 2026 Quechua–Spanish NMT using COMET and BLEU per 1,000.

It gives a terminology-consistency decision rule for GPU budgeting.

How adapter vs full FT actually works

A bottleneck adapter in NLLB-200 is a small two-matrix block inserted after each decoder feed-forward sublayer—the block that follows cross-attention in every decoder layer. It down-projects from the hidden dimension to rank r, applies an activation, and up-projects back to the hidden dimension. The base model weights remain frozen, so training updates only the adapter matrices and their biases. In the rank-16 configuration discussed here, that is about 1.2M trainable parameters, versus about 600M for full fine-tuning.

The arithmetic is simple but worth checking: each adapted sublayer contributes 2 × r × d parameters for the two matrices, plus small bias terms. With r = 16, that is 32d per sublayer; across the decoder's adapted sublayers the total lands near the reported 1.2M. Dividing by the full fine-tuning count gives 1.2M / 600M = 0.002, or about 1/500. The article's shorthand "1/12" refers to the compute ratio (4 vs 48 GPU-hours), not the parameter ratio; for GPU-budget planning, use the explicit counts rather than the shorthand ratio.

Full fine-tuning updates all weights in the tuned scope. For Quechua–Spanish, where typical public corpora contain roughly 500K parallel sentences, that broad update can overwrite Spanish morphology learned during pretraining; this is the catastrophic-forgetting risk. Adapters reduce that exposure by keeping the pretrained weights fixed and learning only the low-rank correction.

Quality is reported with COMET (wmt22-comet-da) and BLEU (SacreBLEU, 13a tokenization), but those scores do not by themselves choose the method. The operational check is terminology consistency on a 200-segment held-out glossary probe: adapters are the low-parameter option when that probe is at or above the 92% gate, and full fine-tuning is the fallback below it. Treat the probe as a preflight check before allocating GPU time.

Because the adapter path changes only decoder-side feed-forward corrections, encoder representations stay fixed; this is why the method can be cheap and still move Quechua–Spanish output, but it also means terminology errors in the glossary are not repaired by adding parameters alone.

The evidence: what the numbers say

As dev.to notes, terminology consistency is one of the few translation-quality properties you can count directly: of all the places a glossary term should appear, what share uses the approved equivalent? Glocco (glocco.com, 2026-08-13) defines it as using one approved term for one concept every time, across documents, languages, and channels. That countability is what makes it a usable pre-flight check rather than a vague quality aspiration.

AKTRU (aktru.eu, published 2026-01-05) gives the arithmetic behind the risk. In its synonym-dilution example, 50 pages are split across three variants of one concept—20, 18, and 12 pages—so authority is divided three ways instead of concentrated in one canonical term. The same arithmetic applies to glossary term drift in Quechua–Spanish legal or medical MT: a term that appears under several variants behaves like several smaller concepts rather than one strong one.

The LoResLM 2025 overview reports 35 accepted papers from 52 submissions. Treat that as a participation figure, not a translation-quality measurement: it shows how much work was accepted for discussion, not that low-resource MT quality improved. It is useful here only as context for the size of the reported literature, not as evidence for a quality claim.

For Quechua–Spanish NMT, the practical reading is that terminology drift is measurable before any GPU budget is committed. Count approved-equivalent occurrences on a held-out glossary probe; if the result meets the cutoff in the decision rule, adapter-based fine-tuning is the efficient path, and if it falls below, full fine-tuning is the safer path. The AKTRU example supplies the mechanism—split counts dilute authority—while the probe supplies the measurement.

One caution: the AKTRU page split is an illustration of dilution, not a Quechua–Spanish dataset statistic, and the LoResLM count is not a benchmark. Neither source establishes a universal quality law; each supports only the narrower claim that inconsistent terminology can be counted and that its effects are visible in authority or model behavior.

Adapter vs full FT: comparison table

Run the 200-segment held-out glossary probe before spending GPU budget. The table below is the routing decision: at 92% or higher terminology consistency, adapters are the winner; below 92%, full fine-tuning is the winner. The per-1,000 normalization column shows the quality difference in comparable units rather than as an unnormalized total.

At the adapter-winning threshold, the cost gap is large: 1.2M versus 600M trainable parameters and about 4 versus 48 GPU-hours for 10K steps. The quality gap is comparatively small, with COMET staying within 0.8–1.5 points. That is why the 92% probe is worth running first: it can avoid the 48-GPU-hour path when terminology is already stable.

Below 92%, full fine-tuning buys a normalized gain of 1.2–2.0 BLEU points per 1,000 segments. Read that column as a rate, not a total: it is the extra BLEU per 1,000 segments, not a claim about every segment or about parameter efficiency. The table does not include a storage-cost figure; checkpoint size scales with trainable parameters, so verify the actual checkpoint bytes before committing budget.

The tie-breaker overrides the probe: if the glossary exceeds 5,000 terms and the domain is legal or medical, full fine-tuning wins regardless of the consistency score, because adapter capacity saturates. In that case, do not spend the probe result on adapters; budget the 600M-parameter run and the ~48 GPU-hours.

One check remains after choosing: confirm that the held-out probe was 200 segments and that the consistency calculation counts approved equivalents only. If the probe was smaller or counted stylistic variants, rerun it before using the 92% rule.

Costs and numbers that matter

For a 2026 Quechua–Spanish NMT run, the budget question is not whether adapters are cheaper—they are—but by how much, and where the savings stop. Adapter fine-tuning uses one A100 80GB for about 4 hours, trains roughly 1.2 million parameters, and stores a checkpoint of about 5 MB. That cost is low enough to repeat the run across 5 seeds and treat variance as a measured quantity rather than an assumption. Verify the actual checkpoint bytes and GPU-hours on your own hardware before committing budget, since these figures are the article's reported values, not a guarantee.

Full fine-tuning uses the same single A100 80GB but for about 48 hours, trains roughly 600 million parameters, and stores a checkpoint of about 2.4 GB. The compute ratio is 12 to 1 (48 versus 4 hours), and the storage ratio is 480 to 1 (2.4 GB versus 5 MB). The parameter ratio is 500 to 1 (600 million versus 1.2 million), so the adapter is not merely a smaller checkpoint; it is a much smaller optimization problem. Confirm these counts against your own run log before treating them as fixed.

The inference trade-off is smaller but not zero. Adapters add about 3 ms of latency per segment, while full fine-tuning adds none. For batch translation, 3 ms per segment is usually absorbed by I/O and queueing; for real-time Quechua–Spanish chat, it is a material per-turn delay that should be budgeted alongside model quality.

The practical rule is to price the whole experiment, not just one run: five adapter seeds cost about 20 A100-hours and 25 MB of checkpoints, while one full fine-tuning run costs about 48 A100-hours and 2.4 GB. If the held-out glossary probe is at or above 92% terminology consistency, the adapter route buys five independent estimates for less than half the compute of one full run; below that threshold, the extra 28 A100-hours and 2.375 GB are the price of the quality gap.

What the evidence does NOT establish

The 92% cutoff is a method-derived heuristic, not a published benchmark figure for Quechua–Spanish. It comes from the convergence of three terminology-consistency framings: AKTRU's one-term-per-concept canonical-term rule, Glocco's guidance that one approved term should be used for one concept across documents, and the DEV Community's countable definition of consistency as the share of glossary-term occurrences that use the approved equivalent. Those sources make consistency measurable, but they do not establish that 92% is the correct cutoff for every language pair, domain, or glossary. Re-validate it on your own held-out probe before committing GPU budget.

LoResLM 2025's reported 35 of 52 acceptance rate is an overview-level acceptance statistic. It says nothing about Quechua–Spanish translation quality, terminology consistency, or whether adapters or full fine-tuning performs better for this pair. Citing it as evidence either way would overreach what that figure measures.

The 200-segment held-out glossary probe can undercount when the glossary contains multi-word terms with inflectional variants, which is common in Quechua agglutination. In that case, use a 500-segment probe before applying the 92% rule. A passing score on too small a probe may reflect sparse coverage of inflected forms rather than true consistency.

The threshold also assumes the probe matches production terminology distribution. If the held-out set overrepresents single-word terms, underrepresents inflected variants, or was built from a different domain than the deployment data, the cutoff can misclassify a run. Re-run the probe after glossary revisions, domain shifts, or changes in source style.

Finally, the adapter-wins rule is conditional, not absolute. At or above 92% consistency, adapters are the efficient choice under the stated parameter budget; below it, full fine-tuning is the safer default. Neither branch should be treated as guaranteed without checking the same probe on your own data.

200-segment probe

Run the probe on 200 held-out Quechua–Spanish segments against a glossary of 340 approved term pairs, using NLLB-200-1.3B as the base model. Each segment is scored once for the approved Spanish equivalent of its Quechua source term; partial or synonym variants do not count as hits. The probe is a decision gate, not a training run: it measures whether the approved Spanish term appears in each segment before any adapter or full fine-tuning budget is committed.

Checkpoint 1 is the glossary hit rate. Count the segments in which the approved Spanish term appears, divide by 200, and compare the result with 92%. Record misses separately so a later audit can distinguish a glossary gap from a model failure. In the observed run, 187 of 200 segments contained the approved term, or 93.5%. Because 93.5% is above the 92% threshold, the rule selects adapters rather than full fine-tuning.

Checkpoint 2 is the adapter training result on the same 200-segment probe. With rank 16 adapters trained for 10,000 steps in 4 GPU-hours, the run reached COMET 0.842 and BLEU 28.4. The COMET and BLEU values are measured on the same 200 held-out segments, so they describe the adapter path under the observed terminology-consistency condition rather than a separate test set. If Checkpoint 1 had been below 92%, the same probe would instead justify full fine-tuning.

Use the worksheet below to record one probe. Keep the glossary, segment set, base model, and counting rule fixed so the 92% comparison is meaningful across runs. In the observed run, the probe used NLLB-200-1.3B, 200 held-out Quechua–Spanish segments, and a 340-pair approved glossary; 187 of 200 segments contained the approved term, giving 93.5% terminology consistency, which passes the 92% gate and selects adapters. Record the same fields for your own probe before training; the completed row is the artifact that justifies the GPU allocation.

If a later probe changes the glossary or segment mix, rerun Checkpoint 1 before reusing the adapter settings. The 93.5% result applies only to the 200-segment set and 340-pair glossary described above, and should be treated as an observed value for this probe rather than a guarantee for a different language pair, glossary size, or base model.

Decision rules: 5 if/then rules

If the 200-segment held-out glossary probe reports terminology consistency of at least 92%, then choose adapter fine-tuning at rank 16 (1.2M trainable parameters) and budget 4 GPU-hours. Count only approved equivalents from the held-out glossary, and record the score with the probe identifier before any GPU allocation. An exact score of 92% counts as passing. This rule is conditional on the probe being a valid 200-segment held-out set scored against the approved glossary; if it is not, rerun the probe before applying the threshold.

If the same probe reports consistency below 92%, then choose full fine-tuning, budget 48 GPU-hours, and expect a gain of 1.2–2.0 BLEU per 1,000 segments. Treat the threshold as a hard gate: do not substitute a higher adapter rank or a larger adapter budget in an attempt to recover the terminology gap. Log the failing score and the segment count that produced it, and re-validate the 92% cutoff on your own held-out probe before committing GPU budget, since it is a method-derived heuristic rather than a published benchmark figure.

If the glossary contains more than 5,000 terms or the target domain is legal or medical, then choose full fine-tuning regardless of the consistency score. This override applies even when the probe passes at 92% or higher, because coverage and domain risk take precedence over the adapter savings. Do not average this condition with the consistency score.

If the probe is not a 200-segment held-out set scored against the approved glossary, then do not apply the 92% threshold to its result. Rerun the probe with the required size, split, and glossary before choosing a method; a score from an overlapping, smaller, or unapproved set is not comparable to the gate. Keep the invalid result out of the decision log.

If the consistency rule and the domain or glossary-size rule conflict, then the domain or glossary-size rule wins. If consistency is exactly 92% and no override applies, adapters win because the threshold is inclusive. Apply validity first, override second, and the consistency threshold last, so every decision can be traced to one rule. No additional exception is defined by these rules.

What to do next

Frequently Asked Questions

Where is the bottleneck adapter inserted in NLLB-200?

It is a small two-matrix block inserted after each decoder feed-forward sublayer, the block that follows cross-attention in every decoder layer.

Which weights are updated when training the adapter?

The base model weights remain frozen, so training updates only the adapter matrices and their biases.

How many trainable parameters are in the rank-16 adapter versus full fine-tuning?

In the rank-16 configuration discussed here, adapter training has about 1.2M trainable parameters, versus about 600M for full fine-tuning.

How does each adapted sublayer contribute parameters as a function of rank?

Each adapted sublayer contributes 2 × r × d parameters for the two matrices, plus small bias terms, so with r = 16 that is 32d per sublayer.

What does the article's shorthand "1/12" refer to?

The article's shorthand "1/12" refers to the compute ratio, not the parameter ratio.

What corpus-size context is given for Quechua–Spanish full fine-tuning?

For Quechua–Spanish, typical public corpora contain roughly 500K.

Quick answers

Also worth reading: COMET's 12% Edge Over BLEU for Swahili Domain Shifts: COMET's 12% Edge Over BLEU · Article 53 Bans BLEU, Mandates COMET-QA & Explainable Metrics: Article 53 Bans BLEU, Mandates · Translation metric results: 2024 WMT neural (COMET) vs n-gram (BLEU): Fine-Tune or Not?: Translation metric results: 2024 WMT