judgeval 0.6.0__py3-none-any.whl → 0.7.1__py3-none-any.whl

judgeval/common/trainer/trainable_model.py

@@ -0,0 +1,238 @@
+from fireworks import LLM
+from .config import TrainerConfig, ModelConfig
+from typing import Optional, Dict, Any, Callable
+from .console import _model_spinner_progress, _print_model_progress
+from judgeval.common.exceptions import JudgmentAPIError
+
+
+class TrainableModel:
+    """
+    A wrapper class for managing model snapshots during training.
+
+    This class automatically handles model snapshot creation and management
+    during the RFT (Reinforcement Fine-Tuning) process,
+    abstracting away manual snapshot management from users.
+    """
+
+    def __init__(self, config: TrainerConfig):
+        """
+        Initialize the TrainableModel.
+
+        Args:
+            config: TrainerConfig instance with model configuration
+        """
+        try:
+            self.config = config
+            self.current_step = 0
+            self._current_model = None
+            self._tracer_wrapper_func = None
+
+            self._base_model = self._create_base_model()
+            self._current_model = self._base_model
+        except Exception as e:
+            raise JudgmentAPIError(
+                f"Failed to initialize TrainableModel: {str(e)}"
+            ) from e
+
+    @classmethod
+    def from_model_config(cls, model_config: ModelConfig) -> "TrainableModel":
+        """
+        Create a TrainableModel from a saved ModelConfig.
+
+        Args:
+            model_config: ModelConfig instance with saved model state
+
+        Returns:
+            TrainableModel instance configured to use the saved model
+        """
+        # Create a TrainerConfig from the ModelConfig
+        trainer_config = TrainerConfig(
+            base_model_name=model_config.base_model_name,
+            deployment_id=model_config.deployment_id,
+            user_id=model_config.user_id,
+            model_id=model_config.model_id,
+            enable_addons=model_config.enable_addons,
+        )
+
+        instance = cls(trainer_config)
+        instance.current_step = model_config.current_step
+
+        if model_config.is_trained and model_config.current_model_name:
+            instance._load_trained_model(model_config.current_model_name)
+
+        return instance
+
+    def _create_base_model(self):
+        """Create and configure the base model."""
+        try:
+            with _model_spinner_progress(
+                "Creating and deploying base model..."
+            ) as update_progress:
+                update_progress("Creating base model instance...")
+                base_model = LLM(
+                    model=self.config.base_model_name,
+                    deployment_type="on-demand",
+                    id=self.config.deployment_id,
+                    enable_addons=self.config.enable_addons,
+                )
+                update_progress("Applying deployment configuration...")
+                base_model.apply()
+            _print_model_progress("Base model deployment ready")
+            return base_model
+        except Exception as e:
+            raise JudgmentAPIError(
+                f"Failed to create and deploy base model '{self.config.base_model_name}': {str(e)}"
+            ) from e
+
+    def _load_trained_model(self, model_name: str):
+        """Load a trained model by name."""
+        try:
+            with _model_spinner_progress(
+                f"Loading and deploying trained model: {model_name}"
+            ) as update_progress:
+                update_progress("Creating trained model instance...")
+                self._current_model = LLM(
+                    model=model_name,
+                    deployment_type="on-demand-lora",
+                    base_id=self.config.deployment_id,
+                )
+                update_progress("Applying deployment configuration...")
+                self._current_model.apply()
+            _print_model_progress("Trained model deployment ready")
+
+            if self._tracer_wrapper_func:
+                self._tracer_wrapper_func(self._current_model)
+        except Exception as e:
+            raise JudgmentAPIError(
+                f"Failed to load and deploy trained model '{model_name}': {str(e)}"
+            ) from e
+
+    def get_current_model(self):
+        return self._current_model
+
+    @property
+    def chat(self):
+        """OpenAI-compatible chat interface."""
+        return self._current_model.chat
+
+    @property
+    def completions(self):
+        """OpenAI-compatible completions interface."""
+        return self._current_model.completions
+
+    def advance_to_next_step(self, step: int):
+        """
+        Advance to the next training step and update the current model snapshot.
+
+        Args:
+            step: The current training step number
+        """
+        try:
+            self.current_step = step
+
+            if step == 0:
+                self._current_model = self._base_model
+            else:
+                model_name = f"accounts/{self.config.user_id}/models/{self.config.model_id}-v{step}"
+                with _model_spinner_progress(
+                    f"Creating and deploying model snapshot: {model_name}"
+                ) as update_progress:
+                    update_progress("Creating model snapshot instance...")
+                    self._current_model = LLM(
+                        model=model_name,
+                        deployment_type="on-demand-lora",
+                        base_id=self.config.deployment_id,
+                    )
+                    update_progress("Applying deployment configuration...")
+                    self._current_model.apply()
+                _print_model_progress("Model snapshot deployment ready")
+
+                if self._tracer_wrapper_func:
+                    self._tracer_wrapper_func(self._current_model)
+        except Exception as e:
+            raise JudgmentAPIError(
+                f"Failed to advance to training step {step}: {str(e)}"
+            ) from e
+
+    def perform_reinforcement_step(self, dataset, step: int):
+        """
+        Perform a reinforcement learning step using the current model.
+
+        Args:
+            dataset: Training dataset for the reinforcement step
+            step: Current step number for output model naming
+
+        Returns:
+            Training job object
+        """
+        try:
+            model_name = f"{self.config.model_id}-v{step + 1}"
+            return self._current_model.reinforcement_step(
+                dataset=dataset,
+                output_model=model_name,
+                epochs=self.config.epochs,
+                learning_rate=self.config.learning_rate,
+                accelerator_count=self.config.accelerator_count,
+                accelerator_type=self.config.accelerator_type,
+            )
+        except Exception as e:
+            raise JudgmentAPIError(
+                f"Failed to start reinforcement learning step {step + 1}: {str(e)}"
+            ) from e
+
+    def get_model_config(
+        self, training_params: Optional[Dict[str, Any]] = None
+    ) -> ModelConfig:
+        """
+        Get the current model configuration for persistence.
+
+        Args:
+            training_params: Optional training parameters to include in config
+
+        Returns:
+            ModelConfig instance with current model state
+        """
+        current_model_name = None
+        is_trained = False
+
+        if self.current_step > 0:
+            current_model_name = f"accounts/{self.config.user_id}/models/{self.config.model_id}-v{self.current_step}"
+            is_trained = True
+
+        return ModelConfig(
+            base_model_name=self.config.base_model_name,
+            deployment_id=self.config.deployment_id,
+            user_id=self.config.user_id,
+            model_id=self.config.model_id,
+            enable_addons=self.config.enable_addons,
+            current_step=self.current_step,
+            total_steps=self.config.num_steps,
+            current_model_name=current_model_name,
+            is_trained=is_trained,
+            training_params=training_params,
+        )
+
+    def save_model_config(
+        self, filepath: str, training_params: Optional[Dict[str, Any]] = None
+    ):
+        """
+        Save the current model configuration to a file.
+
+        Args:
+            filepath: Path to save the configuration file
+            training_params: Optional training parameters to include in config
+        """
+        model_config = self.get_model_config(training_params)
+        model_config.save_to_file(filepath)
+
+    def _register_tracer_wrapper(self, wrapper_func: Callable):
+        """
+        Register a tracer wrapper function to be reapplied when models change.
+
+        This is called internally by the tracer's wrap() function to ensure
+        that new model instances created during training are automatically wrapped.
+
+        Args:
+            wrapper_func: Function that wraps a model instance with tracing
+        """
+        self._tracer_wrapper_func = wrapper_func

judgeval/common/trainer/trainer.py

@@ -0,0 +1,301 @@
+import asyncio
+import time
+from typing import Optional, Callable, Any, List, Union
+from fireworks import Dataset
+from .config import TrainerConfig, ModelConfig
+from .trainable_model import TrainableModel
+from judgeval.tracer import Tracer
+from judgeval.judgment_client import JudgmentClient
+from judgeval.scorers import BaseScorer, APIScorerConfig
+from judgeval.data import Example
+from .console import _spinner_progress, _print_progress, _print_progress_update
+from judgeval.common.exceptions import JudgmentAPIError
+
+
+class JudgmentTrainer:
+    """
+    A reinforcement learning trainer for Judgment models using Fine-Tuning.
+
+    This class handles the iterative training process where models are improved
+    through reinforcement learning fine-tuning based on generated rollouts and rewards.
+    """
+
+    def __init__(
+        self,
+        config: TrainerConfig,
+        trainable_model: TrainableModel,
+        tracer: Tracer,
+        project_name: Optional[str] = None,
+    ):
+        """
+        Initialize the JudgmentTrainer.
+
+        Args:
+            config: TrainerConfig instance with training parameters. If None, uses default config.
+            tracer: Optional tracer for observability
+            trainable_model: Optional trainable model instance
+            project_name: Project name for organizing training runs and evaluations
+        """
+        try:
+            self.config = config
+            self.tracer = tracer
+            self.tracer.show_trace_urls = False
+            self.project_name = project_name or "judgment_training"
+
+            if trainable_model is None:
+                self.trainable_model = TrainableModel(self.config)
+            else:
+                self.trainable_model = trainable_model
+
+            self.judgment_client = JudgmentClient()
+        except Exception as e:
+            raise JudgmentAPIError(
+                f"Failed to initialize JudgmentTrainer: {str(e)}"
+            ) from e
+
+    async def generate_rollouts_and_rewards(
+        self,
+        agent_function: Callable[[Any], Any],
+        scorers: List[Union[APIScorerConfig, BaseScorer]],
+        prompts: List[Any],
+        num_prompts_per_step: Optional[int] = None,
+        num_generations_per_prompt: Optional[int] = None,
+        concurrency: Optional[int] = None,
+    ):
+        """
+        Generate rollouts and compute rewards using the current model snapshot.
+        Each sample contains multiple generations for reinforcement learning optimization.
+
+        Args:
+            agent_function: Function/agent to call for generating responses
+            scorers: List of scorer objects to evaluate responses
+            prompts: List of prompts to use for training
+            num_prompts_per_step: Number of prompts to use per step (defaults to config value, limited by prompts list length)
+            num_generations_per_prompt: Generations per prompt (defaults to config value)
+            concurrency: Concurrency limit (defaults to config value)
+
+        Returns:
+            List of dataset rows containing samples with messages and evaluations
+        """
+        num_prompts_per_step = min(
+            num_prompts_per_step or self.config.num_prompts_per_step, len(prompts)
+        )
+        num_generations_per_prompt = (
+            num_generations_per_prompt or self.config.num_generations_per_prompt
+        )
+        concurrency = concurrency or self.config.concurrency
+
+        semaphore = asyncio.Semaphore(concurrency)
+
+        @self.tracer.observe(span_type="function")
+        async def generate_single_response(prompt_id, generation_id):
+            async with semaphore:
+                prompt_input = prompts[prompt_id]
+                response_data = await agent_function(**prompt_input)
+                messages = response_data.get("messages", [])
+
+                try:
+                    traced_messages = self.tracer.get_current_message_history()
+                    if traced_messages:
+                        messages = traced_messages
+                except Exception as e:
+                    print(f"Warning: Failed to get message history from trace: {e}")
+                    pass
+
+                example = Example(
+                    input=prompt_input,
+                    messages=messages,
+                    actual_output=response_data,
+                )
+
+                scoring_results = self.judgment_client.run_evaluation(
+                    examples=[example],
+                    scorers=scorers,
+                    project_name=self.project_name,
+                    eval_run_name=f"training_step_{self.trainable_model.current_step}_prompt_{prompt_id}_gen_{generation_id}",
+                    show_url=False,
+                )
+
+                if scoring_results and scoring_results[0].scorers_data:
+                    reward = sum(
+                        scorer_data.score
+                        for scorer_data in scoring_results[0].scorers_data
+                    ) / len(scoring_results[0].scorers_data)
+                else:
+                    reward = 0.0
+
+            return {
+                "prompt_id": prompt_id,
+                "generation_id": generation_id,
+                "messages": messages,
+                "evals": {"score": reward},
+            }
+
+        coros = []
+        for prompt_id in range(num_prompts_per_step):
+            for generation_id in range(num_generations_per_prompt):
+                coro = generate_single_response(prompt_id, generation_id)
+                coros.append(coro)
+
+        with _spinner_progress(f"Generating {len(coros)} rollouts..."):
+            num_completed = 0
+            results = []
+
+            for coro in asyncio.as_completed(coros):
+                result = await coro
+                results.append(result)
+                num_completed += 1
+
+        _print_progress(f"Generated {len(results)} rollouts successfully")
+
+        dataset_rows = []
+        for prompt_id in range(num_prompts_per_step):
+            prompt_generations = [r for r in results if r["prompt_id"] == prompt_id]
+            sample_generations = [
+                {"messages": gen["messages"], "evals": gen["evals"]}
+                for gen in prompt_generations
+            ]
+            dataset_rows.append({"samples": sample_generations})
+
+        return dataset_rows
+
+    async def run_reinforcement_learning(
+        self,
+        agent_function: Callable[[Any], Any],
+        scorers: List[Union[APIScorerConfig, BaseScorer]],
+        prompts: List[Any],
+    ) -> ModelConfig:
+        """
+        Run the iterative reinforcement learning fine-tuning loop.
+
+        This method performs multiple steps of reinforcement learning, where each step:
+        1. Advances to the appropriate model snapshot
+        2. Generates rollouts and computes rewards using scorers
+        3. Trains a new model using reinforcement learning
+        4. Waits for training completion
+
+        Args:
+            agent_function: Function/agent to call for generating responses
+            scorers: List of scorer objects to evaluate responses
+            prompts: List of prompts to use for training
+
+        Returns:
+            ModelConfig: Configuration of the trained model for inference and future training
+        """
+
+        _print_progress("Starting reinforcement learning training")
+
+        training_params = {
+            "num_steps": self.config.num_steps,
+            "num_prompts_per_step": self.config.num_prompts_per_step,
+            "num_generations_per_prompt": self.config.num_generations_per_prompt,
+            "epochs": self.config.epochs,
+            "learning_rate": self.config.learning_rate,
+            "accelerator_count": self.config.accelerator_count,
+            "accelerator_type": self.config.accelerator_type,
+            "temperature": self.config.temperature,
+            "max_tokens": self.config.max_tokens,
+        }
+
+        start_step = self.trainable_model.current_step
+
+        for step in range(start_step, self.config.num_steps):
+            step_num = step + 1
+            _print_progress(
+                f"Starting training step {step_num}", step_num, self.config.num_steps
+            )
+
+            self.trainable_model.advance_to_next_step(step)
+
+            dataset_rows = await self.generate_rollouts_and_rewards(
+                agent_function, scorers, prompts
+            )
+
+            with _spinner_progress(
+                "Preparing training dataset", step_num, self.config.num_steps
+            ):
+                dataset = Dataset.from_list(dataset_rows)
+                dataset.sync()
+
+            _print_progress(
+                "Starting reinforcement training", step_num, self.config.num_steps
+            )
+            job = self.trainable_model.perform_reinforcement_step(dataset, step)
+
+            last_state = None
+            with _spinner_progress(
+                "Training job in progress", step_num, self.config.num_steps
+            ):
+                while not job.is_completed:
+                    job.raise_if_bad_state()
+                    current_state = job.state
+
+                    if current_state != last_state:
+                        if current_state in ["uploading", "validating"]:
+                            _print_progress_update(
+                                f"Training job: {current_state} data"
+                            )
+                        elif current_state == "training":
+                            _print_progress_update(
+                                "Training job: model training in progress"
+                            )
+                        else:
+                            _print_progress_update(f"Training job: {current_state}")
+                        last_state = current_state
+
+                    time.sleep(10)
+                    job = job.get()
+                    if job is None:
+                        raise JudgmentAPIError(
+                            "Training job was deleted while waiting for completion"
+                        )
+
+            _print_progress(
+                f"Training completed! New model: {job.output_model}",
+                step_num,
+                self.config.num_steps,
+            )
+
+            dataset.delete()
+
+        _print_progress("All training steps completed!")
+
+        with _spinner_progress("Deploying final trained model"):
+            self.trainable_model.advance_to_next_step(self.config.num_steps)
+
+        return self.trainable_model.get_model_config(training_params)
+
+    async def train(
+        self,
+        agent_function: Callable[[Any], Any],
+        scorers: List[Union[APIScorerConfig, BaseScorer]],
+        prompts: List[Any],
+        rft_provider: Optional[str] = None,
+    ) -> ModelConfig:
+        """
+        Start the reinforcement learning fine-tuning process.
+
+        This is the main entry point for running the reinforcement learning training.
+
+        Args:
+            agent_function: Function/agent to call for generating responses.
+            scorers: List of scorer objects to evaluate responses
+            prompts: List of prompts to use for training
+            rft_provider: RFT provider to use for training. Currently only "fireworks" is supported.
+                         Support for other providers is planned for future releases.
+
+        Returns:
+            ModelConfig: Configuration of the trained model for future loading
+        """
+        try:
+            if rft_provider is not None:
+                self.config.rft_provider = rft_provider
+
+            return await self.run_reinforcement_learning(
+                agent_function, scorers, prompts
+            )
+        except JudgmentAPIError:
+            # Re-raise JudgmentAPIError as-is
+            raise
+        except Exception as e:
+            raise JudgmentAPIError(f"Training process failed: {str(e)}") from e