microsoft--agent-lightning/examples/unsloth/sft_allinone.py

# Copyright (c) Microsoft. All rights reserved.

"""This sample provides an all-in-one script for SFT algorithm.

It's equivalent to running the following commands in parallel:

```bash
agl store
python sft_rollout_runners.py
python sft_algorithm.py
```
"""

from typing import Optional

from math_agent import GsmProblem, load_math_dataset, math_agent
from rich.console import Console
from sft_algorithm import sft_one_iter

from agentlightning import Trainer, setup_logging
from agentlightning.adapter import TraceToTripletBase
from agentlightning.algorithm import Algorithm
from agentlightning.llm_proxy import LLMProxy
from agentlightning.types import Dataset

console = Console()


class UnslothSupervisedFinetuning(Algorithm):
    """Supervised Fine-Tuning (SFT) algorithm implementation using Unsloth.

    This class implements a complete SFT training loop that:
    1. Runs rollouts with the current model
    2. Collects and ranks training data by reward
    3. Fine-tunes the model on top-performing examples
    4. Iterates for multiple rounds of improvement

    Args:
        max_iterations: The maximum number of SFT iterations to perform.
        vllm_port: The port to use for the vLLM inference server.
        train_triplet_fraction: The fraction of top-performing triplets to use for training (0.0 to 1.0).
        initial_model_path: The path to the initial model to start training from.
    """

    def __init__(self, *, max_iterations: int, vllm_port: int, train_triplet_fraction: float, initial_model_path: str):
        # LLM proxy and data adapter are created by the trainer and we can directly use them
        self.max_iterations = max_iterations
        self.vllm_port = vllm_port
        self.train_triplet_fraction = train_triplet_fraction
        self.initial_model_path = initial_model_path

    async def run(
        self, train_dataset: Optional[Dataset[GsmProblem]] = None, val_dataset: Optional[Dataset[GsmProblem]] = None
    ):
        """Execute the SFT training loop. Managed by trainer.

        Args:
            train_dataset: The training dataset of GSM problems to use for rollouts.
            val_dataset: Optional validation dataset (not currently used in SFT).

        Raises:
            ValueError: If train_dataset is None, or required components are missing.
        """
        store = self.get_store()
        llm_proxy = self.get_llm_proxy()
        data_adapter = self.get_adapter()

        # SFT trainer relies on the adapter to convert the trace data to triplets
        if not isinstance(data_adapter, TraceToTripletBase):
            raise ValueError("Data adapter must be a TracerTraceToTriplet.")
        if train_dataset is None:
            raise ValueError("Train dataset must be provided.")
        if val_dataset is not None:
            console.print("[bold red][Algo][/bold red] Validation dataset is not supported in SFT.")
        if llm_proxy is None:
            raise ValueError("LLM proxy must be provided.")

        console.print(f"[bold red][Algo][/bold red] Starting SFT with {self.max_iterations} iterations.")
        console.print(f"[bold red][Algo][/bold red] Initial model path: {self.initial_model_path}")
        model_path = self.initial_model_path
        for iteration in range(self.max_iterations):
            model_path = await sft_one_iter(
                iteration=iteration,
                store=store,
                model_path=model_path,
                train_dataset=train_dataset,
                llm_proxy=llm_proxy,
                data_adapter=data_adapter,
                triplet_fraction=self.train_triplet_fraction,
                vllm_port=self.vllm_port,
            )

        console.print(f"[bold red][Algo][/bold red] Final model path: {model_path}")


if __name__ == "__main__":
    setup_logging()

    algo = UnslothSupervisedFinetuning(
        max_iterations=2,
        vllm_port=12316,
        train_triplet_fraction=0.5,
        initial_model_path="models/version_0",
    )
    trainer = Trainer(
        n_runners=4,
        algorithm=algo,
        llm_proxy=LLMProxy(port=12358),
        # Uncomment the following two lines if you want to rely on proxy-side trace data collection
        # Otherwise, the rollout runner will have an agentops tracer to collect the trace data,
        # and the adapter will be a TracerTraceToTriplet that parses the trace data generated by this tracer
        # adapter=LlmProxyTraceToTriplet(),
        # tracer=OtelTracer(),
    )
    trainer.fit(math_agent, load_math_dataset())