YandexGPT Language Model Fine-Tuning

Fine-Tuning YandexGPT Language Models

YandexGPT is Yandex's language model, available through Yandex Cloud API (Yandex Foundation Models service). YandexGPT fine-tuning is implemented within the service and allows you to adapt the model for specific tasks without managing GPU infrastructure. Key advantage for the Russian market: data is stored in Russian cloud, critical for companies with 152-FZ requirements and corporate security policies.

Available Models and Fine-Tuning Modes

Yandex Cloud provides fine-tuning based on YandexGPT Lite and Pro through Yandex DataSphere service or directly via Foundation Models API. Process is managed through Yandex Cloud Console or CLI.

YandexGPT Lite: fast inference, optimal for classification, structured generation, support chatbots.

YandexGPT Pro: higher quality, complex generation tasks, document analysis, reasoning.

Dataset Format

YandexGPT fine-tuning accepts data in JSON Lines format, where each example is a dialogue with roles:

{
  "request": {
    "messages": [
      {
        "role": "system",
        "text": "You are a bank assistant answering customer questions about products."
      },
      {
        "role": "user",
        "text": "What is the maximum interest rate for the 'Savings Plus' deposit?"
      }
    ]
  },
  "response": "The maximum interest rate for 'Savings Plus' is 16.5% per annum with a 12-month term and sum from 1,000,000 rubles."
}

Recommended volume: 100 to 50,000 examples. Yandex recommends minimum 100 diverse examples for basic adaptation.

Running via Yandex Cloud CLI

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

# Start fine-tuning job
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

Via Python SDK:

import yandexcloud
from yandex.cloud.ai.tuning.v1 import tuning_service_pb2

# Uses gRPC client of Yandex Cloud SDK
# Details in official Yandex Foundation Models documentation

Specifics for Russian Tasks

Legal documents: YandexGPT is trained on significantly larger volume of Russian-language texts, including legislation and judicial practice, compared to most Western models. When fine-tuning on corpus of Russian legislation, baseline quality level is higher.

Financial reporting per Russian standards: specific Russian accounting standards are poorly represented in Western models. YandexGPT is more natural candidate for Russian accounting reporting analysis tasks.

Medical documentation: forms of RF Ministry of Health, medical care standards, clinical guidelines in Russian.

Practical Case: Fine-Tuning for Telecom Operator

Task: automatic support request processing — classification into 28 categories + generation of initial response.

Dataset: 4200 examples from ticket history (real customer requests → category + operator response). Data underwent manual verification and depersonalization.

Result after 5 epochs:

• Classification accuracy: 74% → 91%
• BLEU-4 for responses: 0.21 → 0.54
• Percentage of responses accepted without operator edits: 23% → 67%
• Average request handling time: reduction from 4.2 min to 1.8 min

Comparison with Alternatives

Project Timeline

• Dataset preparation and cleaning: 2–4 weeks
• Training and iterations: 1–2 weeks
• Testing and acceptance: 1 week
• Production integration: 1–2 weeks
• Total: 5–9 weeks