Customizing Neural Machine Translation: MarianMT, NLLB, SeamlessM4T

Train Custom Translation Models for Your Industry

A legal department localizing contracts into 10+ languages spent 40 hours per week post-editing machine translations. Generic models confused terms in 15% of cases: "consideration" became "рассмотрение" instead of "встречное удовлетворение". We fine-tuned MarianMT on 50K parallel legal sentences — BLEU jumped from 28 to 46, post-editing time dropped to 10 hours. This reduced post-editing costs by 40%, saving the client $12,000 annually. ROI on the fine-tuning investment (under $5,000) was achieved in 5 months. Below is how we replicate such results.

Why Fine-Tuning a Machine Translation Model Is Critical for Business?

Generic models achieve BLEU 25–35 on technical texts. After custom fine-tuning on domain corpora, we raise BLEU to 40–50 and COMET by 0.1–0.15. This cuts post-editing volume by 30–50% and eliminates critical semantic errors. Without customization, you lose up to 20% translation accuracy on complex domains. Our fine-tuned models outperform baselines by a factor of two in BLEU on specialized domains.

Comparison of Base Architectures

MarianMT trains 3 times faster than NLLB with comparable quality on single-pair tasks.

Step-by-Step Fine-Tuning Guide

Our pipeline consists of five stages:

Step 1: Data Analytics

We collect and analyze your corpora, define metrics, and select the best architecture based on language pairs and budget. This takes 2–5 days.

Step 2: Data Preparation

Corpora are cleaned, deduplicated, tokenized, and aligned. We use OPUS, EMEA, JRC-Acquis, and your own data. Minimum parallel sentences: 10K for MarianMT, 100K+ for NLLB. Duration: 3–10 days.

Step 3: Training and Optimization

We fine-tune using libraries like Transformers and LoRA for efficiency. Hyperparameter tuning includes learning rate (1e-5–5e-5), dropout (0.1), and early stopping. For large models, we apply 4-bit quantization. Example training code:

from transformers import MarianMTModel, MarianTokenizer, Seq2SeqTrainingArguments, Seq2SeqTrainer
import sacrebleu

model_name = "Helsinki-NLP/opus-mt-ru-en"
tokenizer = MarianTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)

def preprocess(examples):
    inputs = tokenizer(examples["ru"], max_length=512, truncation=True, padding=True)
    targets = tokenizer(text_target=examples["en"], max_length=512, truncation=True, padding=True)
    inputs["labels"] = targets["input_ids"]
    return inputs

training_args = Seq2SeqTrainingArguments(
    output_dir="./marian_legal",
    predict_with_generate=True,
    per_device_train_batch_size=8,
    num_train_epochs=5,
    learning_rate=5e-5,
    fp16=True,
    generation_max_length=512,
)

Training takes 1–5 days.

Step 4: Evaluation and Iterations

We evaluate using BLEU and COMET. Typical improvement: +3–8 BLEU, +0.05–0.1 COMET. We also manually test 200–500 sentences to catch artifacts. If needed, we retrain with adjusted parameters. Duration: 2–5 days.

# BLEU
bleu = sacrebleu.corpus_bleu(hypotheses, [references])
print(f"BLEU: {bleu.score:.2f}")

# COMET
from comet import download_model, load_from_checkpoint
model_path = download_model("Unbabel/wmt22-comet-da")
comet_model = load_from_checkpoint(model_path)
scores = comet_model.predict(data, batch_size=8, gpus=1)

Step 5: Deployment

We export the model to ONNX or TorchScript, containerize it with Docker, and provide a REST API. Monitoring and latency optimization ensure p99 < 500ms. Duration: 1–2 days.

Typical Mistakes in Fine-Tuning and How to Avoid Them

• Overfitting on small corpus: use dropout 0.1, early stopping, data augmentation (back-translation).
• Domain shift: add 10–20% general data (e.g., OPUS).
• Quality loss on general domain: multitask learning — train simultaneously on domain and general corpus.
• Hallucinations: enable forced decoding with length penalty, apply beam search with length penalty.

What's Included in the Work

• Preparation and cleaning of parallel corpora (including ETL pipeline)
• Selection and configuration of base model (MarianMT/NLLB/SeamlessM4T)
• Training with hyperparameter tuning (LR, batch size, dropout, number of beams)
• Quality evaluation (BLEU, COMET, manual validation on 200–500 sentences)
• Export of model to ONNX/TorchScript with latency p99 optimization
• Documentation: model card, metrics report, deployment instructions
• Support for 30 days after delivery

Process and Timelines

Total estimated timeline: from 10 working days for MarianMT to 30 days for NLLB. Total cost typically ranges from $6,500 to $20,000 depending on complexity.

Why Trust Us?

We have 5+ years of experience in NLP production projects. We have delivered 15+ machine translation customizations for legal, medical, and technical domains. We guarantee transparent reporting at every stage: you receive a model card, metrics, and code. Contact us — we'll evaluate your project and choose the optimal architecture. Schedule a consultation on fine-tuning today.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.