Fine-Tuning YandexGPT for Russian Business Tasks

Base YandexGPT makes mistakes in 26% of cases on specific queries—legal wording, financial reports, or internal terminology. Responses become templated, classification inaccurate. Fine-tuning YandexGPT solves this: we adapt the model to your data and scenarios without changing the base architecture. Fine-tuning YandexGPT improves accuracy to 90% and above, confirmed by our cases. All data stays in Russia on Yandex Cloud's certified infrastructure, critical for 152-FZ compliance. We fine-tune YandexGPT for NLP tasks: classification, generation, sentiment analysis. Order a pilot project and see the results on your own data. Our engineers have experience fine-tuning models for banks, telecom, and retail—quality proven by cases with significant savings.

How to prepare a dataset for fine-tuning YandexGPT?

Fine-tuning quality directly depends on the dataset. Yandex recommends at least 100 diverse examples, but in practice the optimal volume is 500 to 5000. The format is JSON Lines, where each example is a dialog with system, user, and assistant roles. Preparation includes:

• Collecting real dialogs from your CRM or chats.
• Cleaning personal data (depersonalization to comply with 152-FZ).
• Labeling correct answers by experts or based on historical data.
• Splitting into training, validation, and test sets (70/15/15).

Example dataset entry:

{
  "request": {
    "messages": [
      {
        "role": "system",
        "text": "You are a bank assistant advising on deposits."
      },
      {
        "role": "user",
        "text": "What is the rate for the 'Savings Plus' deposit with an amount of 500,000?"
      }
    ]
  },
  "response": "For the 'Savings Plus' deposit, the rate is up to 15% per annum for amounts from 500,000 rubles for a 6-month term."
}

What factors affect fine-tuning quality?

Key factors: dataset size, example diversity, number of epochs, and learning rate. We recommend tuning hyperparameters through experimental runs. Typical values:

Launch via Yandex Cloud CLI:

yc ai dataset create \
  --name "bank-faq-dataset" \
  --description "FAQ of banking products" \
  --task-type TextToTextGeneration \
  --upload-format JsonLines \
  --upload-path ./train.jsonl

yc ai tuning create \
  --name "yandexgpt-bank-faq" \
  --base-model-uri "ds://bt1..." \
  --train-datasets uri=<dataset_uri>,weight=1.0 \
  --arguments epochCount=4,learningRate=0.0001,warmupRatio=0.1

Comparison of YandexGPT fine-tuning with alternatives

Fine-tuning YandexGPT is 3x cheaper and 2x faster to deploy than GPT-4o with adaptation to Russian requirements. Comparison with alternatives confirms that for Russian-language tasks, YandexGPT fine-tuning gives the best combination of quality and safety.

Case study: fine-tuning for a telecom operator

Typical mistakes in fine-tuning and how to avoid them

Even with a properly prepared dataset, issues can arise. Main ones:

• Overfitting when epochs exceed 10. Check validation loss every 2 epochs.
• Data drift—after deployment, the model may perform worse due to changed queries. Set up regular monitoring and fine-tune every 1–3 months.
• Incorrect batch size >64 can cause OOM on GPU. Use batch size 16–32 and gradient accumulation if needed.

Stages of work

1. Task and data analysis—we study your datasets, business processes, model requirements. Assess volume, quality, need for augmentation.
2. Dataset preparation—clean, depersonalize, label. Prepare baseline metrics on the original model.
3. Fine-tuning and experiments—run a series of experiments via Yandex DataSphere with different hyperparameters. Select the best model on the validation set.
4. Testing—conduct A/B test on real requests. Evaluate business metrics: accuracy, response time, manual edit rate.
5. Integration and deployment—model is exported to a Yandex Cloud endpoint, connected to your systems via API. Integration with CRM, chats, telephony.
6. Monitoring and re-training—track quality, re-train on new data if needed. Regular updates every 1–3 months.

Guarantee of results

We provide certified engineers experienced in fine-tuning language models. We guarantee transparency at every stage: you receive the dataset, trained model, documentation, and consultation. If quality drops after deployment, we perform correction free of charge for 3 months.

Deliverables

• Ready-to-use dataset for fine-tuning in JSON Lines format.
• Trained model deployed in Yandex Cloud (endpoint).
• Documentation for setup and integration (Swagger, code examples).
• Instructions for monitoring and updating the model.
• Consultation for your engineers (2 hours online).

Estimated timeline and pilot project

Timeline: 3 to 8 weeks depending on task complexity and data volume. Pilot project cost starts at 500,000 rubles. In the first stage, we evaluate your project free of charge: analyze the dataset, select approach, name the cost. Get a consultation — write us in Telegram or leave a request on our website. Order a pilot — contact us to discuss your project.

The concept of fine-tuning describes the basic idea of retraining neural networks.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.