Train Sft

import math, sys
n = int(sys.argv[1])
epochs = max(1, min(10, round(10000 / n)))
batch_size = 2
total_steps = math.ceil(n * epochs / batch_size)
steps_per_epoch = math.ceil(n / batch_size)
warmup_steps = max(10, min(1000, round(steps_per_epoch * 0.05)))
warmup_ratio = round(warmup_steps / total_steps, 4)
print(f"epochs={epochs} batch_size={batch_size} lr=2e-4 schedule=linear warmup_ratio={warmup_ratio}")

    # --- Training complete ---
    step = await model.get_step()
    inference_name = model.get_inference_name()
    client = model.openai_client()

    print("\n" + "=" * 60)
    print("SFT TRAINING COMPLETE")
    print("=" * 60)
    print(f"  Model:          {inference_name}")
    print(f"  Base model:     <BASE_MODEL>")
    print(f"  Training step:  {step}")
    print(f"  Inference URL:  {client.base_url}")
    print(f"  W&B run:        https://wandb.ai/<YOUR_TEAM>/<PROJECT_NAME>/runs/<RUN_NAME>")
    print("=" * 60)

    print("\n--- Python usage (openai SDK) ---\n")
    print(f'''\
from openai import OpenAI

client = OpenAI(
    base_url="{client.base_url}",
    api_key="not-needed",
)

response = client.chat.completions.create(
    model="{inference_name}",
    messages=[
        {{"role": "user", "content": "Your prompt here"}},
    ],
)
print(response.choices[0].message.content)
''')

    print("--- curl usage ---\n")
    print(f'''\
curl {client.base_url}chat/completions \\
  -H "Content-Type: application/json" \\
  -d '{{
    "model": "{inference_name}",
    "messages": [
      {{"role": "user", "content": "Your prompt here"}}
    ]
  }}'
''')

    await backend.close()

from art.local import LocalBackend

backend = LocalBackend()
model = art.TrainableModel(
    name="<RUN_NAME>",
    project="<PROJECT_NAME>",
    base_model="<BASE_MODEL>",
    _internal_config=art.dev.InternalModelConfig(
        engine_args={"gpu_memory_utilization": 0.7},
    ),
)
await model.register(backend)

from art.serverless.backend import ServerlessBackend

backend = ServerlessBackend()  # uses WANDB_API_KEY env var
model = art.TrainableModel(
    name="<RUN_NAME>",
    project="<PROJECT_NAME>",
    base_model="<BASE_MODEL>",
)
await model.register(backend)

"""SFT training script generated by /train-sft wizard."""
import asyncio
import art
<BACKEND_IMPORT>
from art.utils.sft import train_sft_from_file

async def main():
    <BACKEND_SETUP>

    await train_sft_from_file(
        model=model,
        file_path="<FILE_PATH>",
        # Only include these if hyperparameters were computed:
        # epochs=<EPOCHS>,
        # batch_size=<BATCH_SIZE>,
        # peak_lr=<PEAK_LR>,
        # schedule_type="<SCHEDULE_TYPE>",
        # warmup_ratio=<WARMUP_RATIO>,
        verbose=True,
    )

    # ... post-training block + backend.close() ...

if __name__ == "__main__":
    asyncio.run(main())

"""Distillation SFT script generated by /train-sft wizard."""
import asyncio, os
from dotenv import load_dotenv
from openai import AsyncOpenAI
import art
<BACKEND_IMPORT>
from art.utils.sft import create_sft_dataset_iterator

load_dotenv()

async def main():
    teacher_client = AsyncOpenAI(
        api_key=os.environ["<API_KEY_ENV_VAR>"],
        base_url="<TEACHER_API_BASE>",
    )
    prompts = ["<PROMPT_1>", "<PROMPT_2>"]

    trajectories = []
    for prompt in prompts:
        completion = await teacher_client.chat.completions.create(
            model="<TEACHER_MODEL>",
            messages=[{"role": "user", "content": prompt}],
        )
        trajectories.append(
            art.Trajectory(
                messages_and_choices=[
                    {"role": "user", "content": prompt},
                    {"role": "assistant", "content": completion.choices[0].message.content},
                ],
                tools=<TOOLS_OR_NONE>,
                )
        )

    <BACKEND_SETUP>

    for chunk in create_sft_dataset_iterator(
        trajectories,
        epochs=<EPOCHS>,
        batch_size=<BATCH_SIZE>,
        peak_lr=<PEAK_LR>,
        schedule_type="<SCHEDULE_TYPE>",
        warmup_ratio=<WARMUP_RATIO>,
    ):
        await model.train_sft(chunk.trajectories, chunk.config, verbose=True)

    # ... post-training block + backend.close() ...

if __name__ == "__main__":
    asyncio.run(main())