Fine-tuning GPT-4o-mini for SQL generation - results and tips

Sarah ChenMar 22, 2026

I fine-tuned GPT-4o-mini to generate SQL queries for our specific database schema and the results are impressive. Sharing my approach.

Dataset

3,200 (question, SQL) pairs

Covered 15 tables, various JOIN types, aggregations

Included edge cases and complex subqueries

Results

Base GPT-4o-mini: 61% correct SQL on our test set

Fine-tuned: 89% correct SQL

GPT-4o (no fine-tuning): 79% correct SQL

So the fine-tuned mini model outperforms the base GPT-4o at a fraction of the cost!

Key tips

1. Include the schema in every training example's system prompt 2. Add error correction examples (wrong SQL -> correct SQL) 3. Use 2 epochs max — overfitting happens fast 4. Include diverse question phrasings for the same query

6.1k views33 replies88 likes

2 Replies

Karen WuStaffAccepted AnswerMay 23

This is likely related to your network layer, not the API itself. Make sure you're properly handling the SSE stream. Some common causes:

1. Proxy or load balancer buffering chunks 2. Missing Connection: keep-alive header 3. Client-side buffer overflow

Try using the SDK's built-in streaming which handles these edge cases:

with client.chat.completions.create(
    model="gpt-4o",
    messages=[...],
    stream=True
) as stream:
    for text in stream:
        # This handles all edge cases
        pass

Nina Patel May 24

It was our nginx proxy buffering! Added proxy_buffering off; and the issue disappeared. Thanks!