judgeval 0.16.9__py3-none-any.whl → 0.22.2__py3-none-any.whl

judgeval/trainer/trainer.py

@@ -1,405 +1,70 @@
-import asyncio
-import json
-import time
-from typing import Optional, Callable, Any, List, Union, Dict
-from fireworks import Dataset  # type: ignore[import-not-found]
-from .config import TrainerConfig, ModelConfig
+from typing import Optional
+from .config import TrainerConfig
+from .base_trainer import BaseTrainer
+from .fireworks_trainer import FireworksTrainer
 from .trainable_model import TrainableModel
 from judgeval.tracer import Tracer
-from judgeval.tracer.exporters.store import SpanStore
-from judgeval.tracer.exporters import InMemorySpanExporter
-from judgeval.tracer.keys import AttributeKeys
-from judgeval import JudgmentClient
-from judgeval.scorers import ExampleScorer, ExampleAPIScorerConfig
-from judgeval.data import Example
-from .console import _spinner_progress, _print_progress, _print_progress_update
 from judgeval.exceptions import JudgmentRuntimeError
 
 
-class JudgmentTrainer:
+def JudgmentTrainer(
+    config: TrainerConfig,
+    trainable_model: TrainableModel,
+    tracer: Tracer,
+    project_name: Optional[str] = None,
+) -> BaseTrainer:
     """
-    A reinforcement learning trainer for Judgment models using Fine-Tuning.
+    Factory function for creating reinforcement learning trainers.
 
-    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.project_name = project_name or "judgment_training"
-            self.trainable_model = trainable_model
-
-            self.judgment_client = JudgmentClient()
-            self.span_store = SpanStore()
-            self.span_exporter = InMemorySpanExporter(self.span_store)
-        except Exception as e:
-            raise JudgmentRuntimeError(
-                f"Failed to initialize JudgmentTrainer: {str(e)}"
-            ) from e
-
-    def _extract_message_history_from_spans(self) -> List[Dict[str, str]]:
-        """
-        Extract message history from spans in the span store for training purposes.
-
-        This method processes trace spans to reconstruct the conversation flow,
-        extracting messages in chronological order from LLM, user, and tool spans.
-
-        Returns:
-            List of message dictionaries with 'role' and 'content' keys
-        """
-        spans = self.span_store.get_all()
-        if not spans:
-            return []
-
-        messages = []
-        first_found = False
-
-        for span in sorted(spans, key=lambda s: getattr(s, "start_time", 0)):
-            span_attributes = span.attributes or {}
-            span_type = span_attributes.get(AttributeKeys.JUDGMENT_SPAN_KIND, "span")
-
-            if (
-                not span_attributes.get(AttributeKeys.JUDGMENT_OUTPUT)
-                and span_type != "llm"
-            ):
-                continue
-
-            if span_type == "llm":
-                if not first_found and span_attributes.get(
-                    AttributeKeys.JUDGMENT_INPUT
-                ):
-                    input_data: Any = span_attributes.get(
-                        AttributeKeys.JUDGMENT_INPUT, {}
-                    )
-                    if isinstance(input_data, dict) and "messages" in input_data:
-                        input_messages = input_data["messages"]
-                        if input_messages:
-                            first_found = True
-                            for msg in input_messages:
-                                if (
-                                    isinstance(msg, dict)
-                                    and "role" in msg
-                                    and "content" in msg
-                                ):
-                                    messages.append(
-                                        {"role": msg["role"], "content": msg["content"]}
-                                    )
-
-                # Add assistant response from span output
-                output = span_attributes.get(AttributeKeys.JUDGMENT_OUTPUT)
-                if output is not None:
-                    content = str(output)
-                    try:
-                        parsed = json.loads(content)
-                        if isinstance(parsed, dict) and "messages" in parsed:
-                            # Extract the actual assistant message content
-                            for msg in parsed["messages"]:
-                                if (
-                                    isinstance(msg, dict)
-                                    and msg.get("role") == "assistant"
-                                ):
-                                    content = msg.get("content", content)
-                                    break
-                    except (json.JSONDecodeError, KeyError):
-                        pass
-                    messages.append({"role": "assistant", "content": content})
-
-            elif span_type == "user":
-                output = span_attributes.get(AttributeKeys.JUDGMENT_OUTPUT)
-                if output is not None:
-                    content = str(output)
-                    try:
-                        parsed = json.loads(content)
-                        if isinstance(parsed, dict) and "messages" in parsed:
-                            for msg in parsed["messages"]:
-                                if isinstance(msg, dict) and msg.get("role") == "user":
-                                    content = msg.get("content", content)
-                                    break
-                    except (json.JSONDecodeError, KeyError):
-                        pass
-                    messages.append({"role": "user", "content": content})
+    This factory creates and returns provider-specific trainer implementations
+    (FireworksTrainer, VerifiersTrainer, etc.) based on the configured RFT provider.
 
-            elif span_type == "tool":
-                output = span_attributes.get(AttributeKeys.JUDGMENT_OUTPUT)
-                if output is not None:
-                    content = str(output)
-                    try:
-                        parsed = json.loads(content)
-                        if isinstance(parsed, dict) and "messages" in parsed:
-                            for msg in parsed["messages"]:
-                                if isinstance(msg, dict) and msg.get("role") == "user":
-                                    content = msg.get("content", content)
-                                    break
-                    except (json.JSONDecodeError, KeyError):
-                        pass
-                    messages.append({"role": "user", "content": content})
+    The factory pattern allows for easy extension to support multiple training
+    providers without changing the client-facing API.
 
-        return messages
-
-    async def generate_rollouts_and_rewards(
-        self,
-        agent_function: Callable[[Any], Any],
-        scorers: List[Union[ExampleAPIScorerConfig, ExampleScorer]],
-        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
+    Example:
+        config = TrainerConfig(
+            deployment_id="my-deployment",
+            user_id="my-user",
+            model_id="my-model",
+            rft_provider="fireworks"  # or "verifiers" in the future
         )
-        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._extract_message_history_from_spans()
-                    if traced_messages:
-                        messages = traced_messages
-                except Exception as e:
-                    print(f"Warning: Failed to get message history from trace: {e}")
-                    pass
-
-                finally:
-                    self.span_store.spans = []
-
-                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}",
-                )
-
-                if scoring_results and scoring_results[0].scorers_data:
-                    scores = [
-                        scorer_data.score
-                        for scorer_data in scoring_results[0].scorers_data
-                        if scorer_data.score is not None
-                    ]
-                    reward = sum(scores) / len(scores) if scores else 0.0
-                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 = []
+        # User creates and configures the trainable model
+        trainable_model = TrainableModel(config)
+        tracer = Tracer()
 
-            for coro in asyncio.as_completed(coros):
-                result = await coro
-                results.append(result)
-                num_completed += 1
+        # JudgmentTrainer automatically creates the appropriate provider-specific trainer
+        trainer = JudgmentTrainer(config, trainable_model, tracer)
 
-        _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[ExampleAPIScorerConfig, ExampleScorer]],
-        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
+        # The returned trainer implements the BaseTrainer interface
+        model_config = await trainer.train(agent_function, scorers, prompts)
 
         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
+        config: TrainerConfig instance with training parameters including rft_provider
+        trainable_model: Provider-specific trainable model instance (e.g., TrainableModel for Fireworks)
+        tracer: Tracer for observability
+        project_name: Project name for organizing training runs and evaluations
 
         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 JudgmentRuntimeError(
-                            "Training job was deleted while waiting for completion"
-                        )
+        Provider-specific trainer instance (FireworksTrainer, etc.) that implements
+        the BaseTrainer interface
 
-            _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[ExampleAPIScorerConfig, ExampleScorer]],
-        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 JudgmentRuntimeError:
-            # Re-raise JudgmentAPIError as-is
-            raise
-        except Exception as e:
-            raise JudgmentRuntimeError(f"Training process failed: {str(e)}") from e
+    Raises:
+        JudgmentRuntimeError: If the specified provider is not supported
+    """
+    provider = config.rft_provider.lower()
+
+    if provider == "fireworks":
+        return FireworksTrainer(config, trainable_model, tracer, project_name)
+    elif provider == "verifiers":
+        # Placeholder for future implementation
+        raise JudgmentRuntimeError(
+            "Verifiers provider is not yet implemented. "
+            "Currently supported providers: 'fireworks'"
+        )
+    else:
+        raise JudgmentRuntimeError(
+            f"Unsupported RFT provider: '{config.rft_provider}'. "
+            f"Currently supported providers: 'fireworks'"
+        )

judgeval/utils/guards.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 from typing import TYPE_CHECKING
+from judgeval.logger import judgeval_logger
 
 if TYPE_CHECKING:
     from typing import TypeVar
@@ -8,24 +9,27 @@ if TYPE_CHECKING:
     T = TypeVar("T")
 
 
-def expect_exists(value: T | None, message: str) -> T:
-    if value is None:
-        raise ValueError(message)
+def expect_exists(value: T | None, message: str, default: T) -> T:
+    if not value:
+        judgeval_logger.error(message)
+        return default
 
     return value
 
 
-def expect_api_key(api_key: str | None) -> str:
+def expect_api_key(api_key: str | None) -> str | None:
     return expect_exists(
         api_key,
         "API Key is not set, please set JUDGMENT_API_KEY in the environment variables or pass it as `api_key`",
+        default=None,
     )
 
 
-def expect_organization_id(organization_id: str | None) -> str:
+def expect_organization_id(organization_id: str | None) -> str | None:
     return expect_exists(
         organization_id,
         "Organization ID is not set, please set JUDGMENT_ORG_ID in the environment variables or pass it as `organization_id`",
+        default=None,
     )
 
 

judgeval/utils/project.py

@@ -0,0 +1,15 @@
+from judgeval.utils.decorators.dont_throw import dont_throw
+import functools
+from judgeval.api import JudgmentSyncClient
+
+
+@dont_throw
+@functools.lru_cache(maxsize=64)
+def _resolve_project_id(project_name: str, api_key: str, organization_id: str) -> str:
+    """Resolve project_id from project_name using the API."""
+    client = JudgmentSyncClient(
+        api_key=api_key,
+        organization_id=organization_id,
+    )
+    response = client.projects_resolve({"project_name": project_name})
+    return response["project_id"]

judgeval/utils/serialize.py

@@ -247,7 +247,7 @@ encoders_by_class_tuples = generate_encoders_by_class_tuples(ENCODERS_BY_TYPE)
 # Seralize arbitrary object to a json string
 def safe_serialize(obj: Any) -> str:
     try:
-        return orjson.dumps(json_encoder(obj)).decode()
+        return orjson.dumps(json_encoder(obj), option=orjson.OPT_NON_STR_KEYS).decode()
     except Exception as e:
         judgeval_logger.warning(f"Error serializing object: {e}")
-        return orjson.dumps(repr(obj)).decode()
+        return repr(obj)

judgeval/version.py

@@ -1,4 +1,4 @@
-__version__ = "0.16.9"
+__version__ = "0.22.2"
 
 
 def get_version() -> str:

{judgeval-0.16.9.dist-info → judgeval-0.22.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: judgeval
-Version: 0.16.9
+Version: 0.22.2
 Summary: Judgeval Package
 Project-URL: Homepage, https://github.com/JudgmentLabs/judgeval
 Project-URL: Issues, https://github.com/JudgmentLabs/judgeval/issues
@@ -63,8 +63,7 @@ Judgeval's agent monitoring infra provides a simple harness for integrating GRPO
 await trainer.train(
     agent_function=your_agent_function,  # entry point to your agent
     scorers=[RewardScorer()],  # Custom scorer you define based on task criteria, acts as reward
-    prompts=training_prompts,  # Tasks
-    rft_provider="fireworks"
+    prompts=training_prompts  # Tasks
 )
 ```