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

judgeval/trainer/fireworks_trainer.py

@@ -0,0 +1,396 @@
+import asyncio
+import json
+from typing import Optional, Callable, Any, List, Union, Dict
+from fireworks import Dataset  # type: ignore[import-not-found]
+from .config import TrainerConfig, ModelConfig
+from .base_trainer import BaseTrainer
+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 FireworksTrainer(BaseTrainer):
+    """
+    Fireworks AI implementation of the training provider.
+
+    This trainer uses Fireworks AI's infrastructure for reinforcement learning
+    fine-tuning (RFT) of language models.
+    """
+
+    def __init__(
+        self,
+        config: TrainerConfig,
+        trainable_model: TrainableModel,
+        tracer: Tracer,
+        project_name: Optional[str] = None,
+    ):
+        """
+        Initialize the FireworksTrainer.
+
+        Args:
+            config: TrainerConfig instance with training parameters
+            trainable_model: TrainableModel instance for Fireworks training
+            tracer: Tracer for observability
+            project_name: Project name for organizing training runs and evaluations
+        """
+        try:
+            super().__init__(config, trainable_model, tracer, project_name)
+
+            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 FireworksTrainer: {str(e)}"
+            ) from e
+
+    def _extract_message_history_from_spans(
+        self, trace_id: str
+    ) -> 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.
+
+        Args:
+            trace_id: The trace ID (32-char hex string) to extract message history from
+
+        Returns:
+            List of message dictionaries with 'role' and 'content' keys
+        """
+        spans = self.span_store.get_by_trace_id(trace_id)
+        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 in ("user", "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})
+
+        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
+        )
+        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", [])
+
+                current_span = self.tracer.get_current_span()
+                trace_id = None
+                if current_span and current_span.is_recording():
+                    # Convert trace_id to hex string per OTEL spec
+                    trace_id = format(current_span.get_span_context().trace_id, "032x")
+
+                try:
+                    if trace_id is not None:
+                        traced_messages = self._extract_message_history_from_spans(
+                            trace_id
+                        )
+                        if traced_messages:
+                            messages = traced_messages
+                except Exception as e:
+                    print(f"Warning: Failed to get message history from trace: {e}")
+                    pass
+
+                finally:
+                    if trace_id is not None:
+                        self.span_store.clear_trace(trace_id)
+
+                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 = []
+
+            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[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
+
+        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
+
+                    await asyncio.sleep(10)
+                    job = job.get()
+                    if job is None:
+                        raise JudgmentRuntimeError(
+                            "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[ExampleAPIScorerConfig, ExampleScorer]],
+        prompts: List[Any],
+    ) -> 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
+
+        Returns:
+            ModelConfig: Configuration of the trained model for future loading
+        """
+        try:
+            return await self.run_reinforcement_learning(
+                agent_function, scorers, prompts
+            )
+        except JudgmentRuntimeError:
+            # Re-raise JudgmentRuntimeError as-is
+            raise
+        except Exception as e:
+            raise JudgmentRuntimeError(f"Training process failed: {str(e)}") from e