Fine-tuning чат-моделей: полный цикл

Подготовка тренировочных данных

import json, pandas as pd
from openai import OpenAI

client = OpenAI()

recipe_df = pd.read_csv("data/cookbook_recipes_nlg_10k.csv")
system_msg = "You are a helpful recipe assistant. Extract generic ingredients."

def prepare_example(row):
    return {
        "messages": [
            {"role": "system", "content": system_msg},
            {"role": "user",
             "content": f"Title: {row['title']}\n\nIngredients: {row['ingredients']}\n\nGeneric ingredients: "},
            {"role": "assistant", "content": row["NER"]},
        ]
    }

training_data = recipe_df.loc[0:100].apply(prepare_example, axis=1).tolist()
validation_data = recipe_df.loc[101:200].apply(prepare_example, axis=1).tolist()

def write_jsonl(data, path):
    with open(path, "w") as f:
        for d in data:
            f.write(json.dumps(d) + "\n")

write_jsonl(training_data, "train.jsonl")
write_jsonl(validation_data, "val.jsonl")

Загрузка файлов и запуск job

def upload_file(path):
    with open(path, "rb") as f:
        return client.files.create(file=f, purpose="fine-tune").id

train_id = upload_file("train.jsonl")
val_id   = upload_file("val.jsonl")

job = client.fine_tuning.jobs.create(
    training_file=train_id,
    validation_file=val_id,
    model="gpt-4o-mini-2024-07-18",
    suffix="recipe-ner",
)
print("Job ID:", job.id, "| Status:", job.status)

Мониторинг и инференс

import time

while True:
    status = client.fine_tuning.jobs.retrieve(job.id).status
    print(f"Status: {status}")
    if status in ("succeeded", "failed", "cancelled"):
        break
    time.sleep(60)

ft_model = client.fine_tuning.jobs.retrieve(job.id).fine_tuned_model

# Инференс
resp = client.chat.completions.create(
    model=ft_model,
    messages=[
        {"role": "system", "content": system_msg},
        {"role": "user", "content": "Title: Pasta\nIngredients: [...]\nGeneric ingredients: "},
    ],
    temperature=0,
    max_tokens=200,
)
print(resp.choices[0].message.content)

Начинайте с 50–100 примеров. Перфоманс растёт линейно с объёмом датасета, поэтому итеративно увеличивайте размер, пока не достигнете целевой точности.