Adapting T5, BART, and ruT5 for Domain-Specific Text Summarization

General-purpose models like GPT-4o handle news summaries well, but on a legal contract they miss key clauses—force majeure or notice period. Our practical experience shows that domain adaptation dramatically changes quality. Fine-tuning on domain-specific data teaches the model to highlight what matters in your context, be it medical records, court rulings, or technical documentation. Our engineers adapt architectures to task specifics: style, terminology, and structure. The result is reliable summarization. Inference cost savings compared to commercial LLMs can reach 10x; for example, fine-tuning ruT5 on 3000 pairs costs about $500 and reduces monthly inference costs by 70%. Contact us for a free project assessment.

What Problems Does Fine-tuning Solve?

Style Mismatch. Generative models tend toward conversational phrasing. Business reports need formal language, news requires brevity, and medical texts demand precise terminology. Fine-tuning adapts the style to the domain, leveraging transfer learning.

Factual Hallucinations. Models may insert non-existent figures or names, critical in legal and financial contexts. Domain-specific fine-tuning reduces hallucination rates by 30–50%. Research indicates that adapted models produce 20% fewer fabricated details.

Long Context. Documents like court rulings exceed 8K tokens. T5 and BART with 1024-token windows require chunking—we use a sliding window strategy with overlap to preserve context.

How We Adapted the Model for Legal Documents

A client—a legal department at a large company—needed to summarize supply contracts. Standard models (GPT-4o) produced paraphrases rather than extracts, missing dates, amounts, and termination conditions.

Our approach:

• Collected 2000 contract-summary pairs from a 5-year archive.
• Used ruT5-base with LoRA (rank 16, target modules q_proj, v_proj).
• Trained on one A100 80GB, 3 epochs, batch size 8, gradient accumulation 2. Loss converged in 2 hours.
• After fine-tuning, ROUGE-1 increased from 0.32 to 0.48, BERTScore from 0.78 to 0.89.
• Human evaluation: accuracy 4.2/5, completeness 4.5/5.

Result: The model extracts all essential conditions with hallucinations below 2%. This demonstrates that fine-tuning T5 for domain-specific summarization yields ROUGE scores 1.8x higher than the base model.

Architecture Comparison for Summarization

For Russian domains, we recommend ruT5 or ruBART—they are compact and achieve strong performance with 2000+ examples. After fine-tuning, these models retain document structure 1.5x better than multilingual counterparts. Choosing the right architecture yields significant cost savings.

Why Fine-tune on Domain-specific Data?

Training on a general corpus like Common Crawl does not capture your business specifics. Terms, document structure, and field importance require adjustment. Fine-tuning for domain adaptation teaches the model that "material terms" in a contract differ from "weather" in news. In practice, an adapted model achieves ROUGE-L 15–20 points higher than the base model.

Data Requirements for Quality Fine-tuning

Our experience shows that 1000–5000 pairs are enough to gain 15–20 ROUGE points. With less data, we use augmentation: back-translation, synonym replacement, and recursive summarization via GPT-4o. Synthetic data cannot replace real data—always mix them in a 1:3 ratio. Evaluation uses BERTScore alongside ROUGE. Domain adaptation NLP is key for transferring knowledge to new domains.

Process

1. Domain and requirements analysis—collect a representative sample, define style and summary length.
2. Data preparation—cleaning, labeling (if pairs missing), chunking for long documents using concatenation with a separator and sliding window.
3. Architecture selection and fine-tuning—experiment with T5, BART, sometimes mT5. Tune hyperparameters: learning rate (1e-4…5e-5), batch size, epochs. Use EarlyStopping from Hugging Face.
4. Evaluation—ROUGE + BERTScore + human evaluation. If quality is below threshold, revisit data or architecture.
5. Deployment—serve via Triton Inference Server or FastAPI with p99 latency optimized to 500ms. Use INT8 quantization for GPU acceleration.

What’s Included

• Analysis of your data and domain with architecture selection.
• Preparation and augmentation of the training set.
• Fine-tuning with hyperparameter optimization.
• Evaluation via ROUGE, BERTScore, and human testing.
• Model deployment (Triton, FastAPI) with documentation.
• Training your team to work with the model.
• Stability guarantee—additional epochs if needed.
• Deliverables include model weights, inference scripts, API documentation, and a support period of 3 months.

Contact us for a consultation: we will help assess your project, select architecture, and estimate the budget. We bring 7+ years of experience and 30+ NLP and summarization projects. Request a preliminary analysis—we will demonstrate the quality improvement fine-tuning can achieve on your data.

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.