freesolo 0.2.13

Tracing, evaluation, and training utilities for LLM applications.

freesolo

freesolo is the Python SDK used by Freesolo-generated training repos.

The SDK gives generated repos one shared surface for:

• loading the approved training contract
• loading datasets and building training conversations
• defining the repo-specific task environment
• running contract-aligned evaluations
• running GEPA prompt optimization
• launching SFT and GRPO training
• optionally exporting OpenTelemetry traces

The main idea is that a generated repo should contain only the task-specific files under freesolo/, while the reusable training, evaluation, dataset, contract, and tracing behavior comes from this package.

Install

pip install freesolo

From a checkout:

cd freesolo-sdk
export PYTHONPATH="$PWD/pypi"

Credentials

Most workflows that upload results or start hosted work need a Freesolo API key:

export FREESOLO_API_KEY=fslo_...

Optional environment variables:

• FREESOLO_BASE_URL: defaults to https://api.freesolo.co
• OPENROUTER_API_KEY: hosted LLM-as-judge scorers
• TINKER_API_KEY: SFT and GRPO training
• WANDB_API_KEY: experiment tracking when enabled by the generated repo

Generated Repo Flow

The SDK is built around the files that Freesolo agents generate in a target repo:

freesolo/TRAINING_CONTRACT.md
freesolo/config.py
freesolo/environment.py
freesolo/data.py
freesolo/eval.py
freesolo/gepa.py
freesolo/training.py

A normal generated repo flow is:

1. Write or approve freesolo/TRAINING_CONTRACT.md.
2. Define the task once in freesolo/environment.py.
3. Run evals against candidate model outputs with the same environment and contract.
4. Use the same environment for GEPA, SFT, and GRPO.
5. Add tracing only when you need observability for app or SDK spans.

Tracing is not the center of the SDK. It is optional instrumentation around the contract/eval/training loop.

Environment

Environment is the task adapter. It defines how examples become model prompts and how model responses are scored.

from freesolo.datasets import TaskExample
from freesolo.environments import Environment, RewardResult


class RepoEnvironment(Environment):
    def build_prompt_messages(self, example: TaskExample, prompt_text: str):
        return [
            {"role": "system", "content": prompt_text},
            {"role": "user", "content": example.task},
        ]

    def score_response(self, example: TaskExample, response_text: str) -> RewardResult:
        expected = str(example.expected_output or "").strip()
        actual = response_text.strip()
        passed = actual == expected
        return RewardResult(
            name="exact_match",
            score=1.0 if passed else 0.0,
            success=passed,
            threshold=1.0,
            reason="matched expected output" if passed else "mismatch",
            return_type="binary",
        )


def load_environment(**_: object) -> Environment:
    return RepoEnvironment()

Generated repo helpers should pass this reference through SDK APIs:

ENVIRONMENT_REFERENCE = "freesolo/environment.py:load_environment"

That keeps evals, GEPA, SFT, and GRPO aligned on one prompt and reward definition.

Evaluations

Environment evals run model outputs through the contract and environment reward logic, then upload the result to Freesolo.

from openai import OpenAI

from freesolo.datasets import TaskExample
from freesolo.environments import EnvironmentGeneration
from freesolo.evaluation import EvaluationClient

from config import CONTRACT_PATH, ENVIRONMENT_REFERENCE


client = OpenAI()


def generate(messages: list[dict[str, str]], example: TaskExample):
    response = client.chat.completions.create(
        model="gpt-4.1-mini",
        messages=messages,
    )
    return EnvironmentGeneration(
        response_text=response.choices[0].message.content or "",
        total_tokens=response.usage.total_tokens if response.usage else None,
    )


results = EvaluationClient().run_environment(
    name="dev-eval",
    source="runs/eval/dev.jsonl",
    contract_path=CONTRACT_PATH,
    environment=ENVIRONMENT_REFERENCE,
    generate=generate,
)

For smaller scripts and CI checks, custom scorers are also supported:

from typing import Any

from freesolo.evaluation import BinaryResponse, CustomScorer, EvaluationClient


class NoEmptyAnswer(CustomScorer[BinaryResponse]):
    async def score(self, row: dict[str, Any]) -> BinaryResponse:
        ok = bool(str(row.get("actual_output", "")).strip())
        return BinaryResponse(value=ok, reason="actual_output is non-empty")


results = EvaluationClient().run(
    name="non-empty-answer",
    data=[{"actual_output": "hello"}],
    scorers=[NoEmptyAnswer()],
)

GEPA And Training

GEPA, SFT, and GRPO use the same contract, datasets, and environment adapter as evals. Generated repos should call the SDK helpers rather than copying trainer or optimizer internals.

from freesolo.training import train_grpo, train_sft

from config import (
    BASE_MODEL,
    CONTRACT_PATH,
    ENVIRONMENT_REFERENCE,
    GRPO_DATASET_PATH,
    GRPO_LOG_DIR,
    SFT_CONFIG,
    SFT_DATASET_PATH,
    SFT_LOG_DIR,
)


def run_sft() -> int:
    return train_sft(
        contract_path=CONTRACT_PATH,
        dataset_path=SFT_DATASET_PATH,
        environment=ENVIRONMENT_REFERENCE,
        log_dir=SFT_LOG_DIR,
        base_model=BASE_MODEL,
        sft_config=SFT_CONFIG,
    )


def run_grpo() -> int:
    return train_grpo(
        contract_path=CONTRACT_PATH,
        dataset_path=GRPO_DATASET_PATH,
        environment=ENVIRONMENT_REFERENCE,
        log_dir=GRPO_LOG_DIR,
        sft_log_dir=SFT_LOG_DIR,
        base_model=BASE_MODEL,
    )

Tracing

Tracing is available for applications or generated repo commands that need span export. Configure it at process startup, then use normal OpenTelemetry spans.

from freesolo.tracing import configure_tracer, force_flush, get_tracer

configure_tracer(service_name="my-training-repo")
tracer = get_tracer()

with tracer.start_as_current_span("eval.batch") as span:
    span.set_attribute("freesolo.dataset", "runs/eval/dev.jsonl")

force_flush()

Runnable Examples

Copy-pasteable examples live in examples/:

• environment.py: task environment used by evals, training, and GEPA.
• support_dataset.py: dataset loading helpers for evals, SFT, GRPO, and GEPA.
• evaluation_from_files.py: run an environment eval from concrete files.
• evaluation_custom_scorer.py: run local custom scorers.
• gepa_prompt_example.py: run the Freesolo GEPA adapter.
• training_sft_grpo.py: start SFT or GRPO training from package APIs.
• tracing_manual_span.py: send one OpenTelemetry span.

Example:

uv run python examples/evaluation_custom_scorer.py --local

Public API

The root freesolo module intentionally exports no functions. Import from the subpackages below; lower-level modules may be importable, but they are implementation helpers unless they appear here or in an example.

Import	Use case
freesolo.contracts.load_contract_text, extract_contract_spec, load_contract_spec, build_oracle_messages	Read contract markdown and build oracle prompt messages.
freesolo.datasets.TaskExample, Dataset, load_dataset	Load task examples and construct labeled conversations for evals or training.
freesolo.environments.Environment, RewardResult, RewardMetric, EnvironmentGeneration	Define task prompt and reward behavior once for evals, GEPA, SFT, and GRPO.
freesolo.evaluation.EvaluationClient	Run custom-scorer evals or environment evals and upload results to Freesolo.
freesolo.evaluation.run_local_evaluation	Run custom scorers locally without uploading results.
freesolo.evaluation.CustomScorer, BinaryResponse, NumericResponse	Define local scorer logic for eval rows.
freesolo.evaluation.HostedJudgeClient and hosted scorer classes	Use hosted LLM-as-judge scorers with OpenRouter-compatible credentials.
freesolo.gepa.GEPASetup, GEPAConfig, DefaultReflectionAgent, attach_gepa, optimize_gepa	Optimize prompts through the GEPA adapter using the same environment and dataset abstractions.
freesolo.training.SftConfig, GrpoConfig, TrainGrpoOptions, train_sft, train_grpo	Start SFT or GRPO training from package APIs.
freesolo.tracing.configure_tracer, get_tracer, force_flush, shutdown	Export OpenTelemetry traces when observability is needed.
freesolo.utils.oracle.generate_ground_truth_records	Generate ground-truth JSONL records from source examples using a contract, environment, and oracle model.
freesolo.utils.upload.upload_tinker_checkpoint_to_huggingface	Upload a Tinker checkpoint to a private Hugging Face model repo.

Package Docs

The generated-repo-facing package notes live next to the modules:

• pypi/freesolo/README.md
• pypi/freesolo/contracts/README.md
• pypi/freesolo/datasets/README.md
• pypi/freesolo/environments/README.md
• pypi/freesolo/evaluation/README.md
• pypi/freesolo/gepa/README.md
• pypi/freesolo/training/README.md
• pypi/freesolo/tracing/README.md