GigaChat (Sber) Language Model Fine-Tuning

Fine-Tuning GigaChat Language Model (Sber)

GigaChat is Sber's language model, available through GigaChat API. Model fine-tuning is implemented within the corporate segment (GigaChat Pro / GigaChat Enterprise) and allows you to adapt the model to your business specifics. The model is originally oriented toward Russian-language market, making it a priority choice for tasks involving Russian content, Russian accounting documents (RSBU), and interactions within Russian legal framework.

Access to GigaChat Fine-Tuning

GigaChat API is provided through Sber Cloud platform. For fine-tuning, an enterprise contract and access to appropriate tier are required. Standard API access provides base models GigaChat Lite, GigaChat Pro, and GigaChat Max; fine-tuning is available through GigaChat Enterprise or special request to corporate department.

Authentication via OAuth 2.0:

import requests
import base64

credentials = base64.b64encode(
    f"{client_id}:{client_secret}".encode()
).decode()

response = requests.post(
    "https://ngw.devices.sberbank.ru:9443/api/v2/oauth",
    headers={
        "Authorization": f"Basic {credentials}",
        "RqUID": "unique-request-id",
    },
    data={"scope": "GIGACHAT_API_CORP"}
)
access_token = response.json()["access_token"]

Data Format for Fine-Tuning

GigaChat accepts training data as dialogue pairs in JSON format. Each example is a complete conversation context with roles:

{
  "messages": [
    {
      "role": "system",
      "content": "You are an insurance company assistant. Help customers understand insurance product conditions."
    },
    {
      "role": "user",
      "content": "What is covered under CASCO insurance for accidents caused by third parties?"
    },
    {
      "role": "assistant",
      "content": "For accidents caused by third parties under CASCO: collision damage regardless of fault is covered, restoration repair costs are covered..."
    }
  ]
}

Advantages for Russian Market

Financial regulatory compliance: GigaChat is trained on Russian regulatory documents, including Central Bank guidelines, RSBU standards, tax reporting forms. This reduces data volume needed for financial task fine-tuning.

Medical documentation: Ministry of Health templates, clinical guidelines, ICD-10, Social Fund integration. Native support for Russian medical terminology.

Government services: work with SMEV documents, department formats, specific public sector terminology.

Data security: data remains within Sber Cloud perimeter, critical for banks, insurance companies, government organizations.

Practical Example: Assistant for Banking Chatbot

Task: fine-tune GigaChat Pro for processing incoming requests to retail bank — answering product questions, routing complex requests to operators.

Dataset: 3500 dialogues from real correspondence (anonymized), covering 45 topics (loans, deposits, cards, transfers, transaction disputes).

Data preparation stages:

1. Extract dialogues from CRM
2. Depersonalize (replace names, card numbers, phones)
3. Filter dialogues with negative outcomes (customer got no answer)
4. Mark complex cases requiring routing
5. Balance by topics (no more than 15% from one category)

Results:

• CSAT (customer satisfaction with bot answer): 3.2 → 4.1 out of 5
• Correct routing percentage: 71% → 94%
• Escalation rate reduction (requests passed to operator): 61% → 38%
• Average dialogue time: -22%

GigaChat Fine-Tuning Limitations

• Closed weights: like GPT-4o, you get hosted endpoint without weight access
• Infrastructure lock-in: only Sber Cloud, no on-premise deployment
• Corporate threshold: fine-tuning unavailable on free tiers
• Context size: 32K tokens — smaller than Qwen2.5 or Claude 3.5 Sonnet

Comparison with Related Solutions

Project Timeline

• Task audit, dataset evaluation: 3–5 days
• Data preparation and depersonalization: 2–4 weeks
• Iterative training: 1–2 weeks
• Testing, A/B: 1 week
• Integration, monitoring: 1–2 weeks
• Total: 5–9 weeks