Recursive-SWE-bench / core /recursive_task.py
recursivelabs's picture
Upload 7 files
e8a0a6a verified
# recursive_swe_bench/core/recursive_task.py
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple, Union
from enum import Enum
import datetime
import uuid
import json
import copy
class TaskStatus(Enum):
"""Status of a recursive task."""
INITIALIZED = "initialized"
IN_PROGRESS = "in_progress"
CONVERGED = "converged"
MAX_ITERATIONS = "max_iterations"
PERFECT_SOLUTION = "perfect_solution"
ABANDONED = "abandoned"
@dataclass
class ProblemState:
"""Represents the current state of a problem in the recursive task."""
problem_id: str
description: str
code_context: Dict[str, Any]
requirements: List[Dict[str, Any]]
difficulty: float # 0.0 to 1.0
evolution_stage: int # How many times the problem has evolved
adaptation_vector: List[float] # Directs how the problem should evolve
@dataclass
class EvaluationResult:
"""Results from evaluating a solution."""
success: bool
score: float # 0.0 to 1.0
execution_results: Dict[str, Any]
error_details: Optional[Dict[str, Any]] = None
test_results: Optional[Dict[str, Any]] = None
metrics: Optional[Dict[str, float]] = None
@dataclass
class Feedback:
"""Structured feedback on a solution."""
summary: str
issues: List[Dict[str, Any]]
suggestions: List[Dict[str, Any]]
focus_areas: List[str]
adaptation_hints: List[Dict[str, Any]]
class ConvergenceCriteria:
"""Criteria for determining when a recursive task has converged."""
def __init__(self, config: Dict[str, Any] = None):
self.config = config or {}
self.score_threshold = self.config.get("score_threshold", 0.95)
self.min_iterations = self.config.get("min_iterations", 1)
self.max_iterations = self.config.get("max_iterations", 10)
self.score_delta_threshold = self.config.get("score_delta_threshold", 0.01)
self.consecutive_plateau_limit = self.config.get("consecutive_plateau_limit", 3)
def has_converged(self, trajectory: "Trajectory") -> bool:
"""Determine if the task has converged based on the trajectory."""
if len(trajectory.steps) < self.min_iterations:
return False
if len(trajectory.steps) >= self.max_iterations:
return True
# Check if we've reached the score threshold
latest_score = trajectory.steps[-1].result.score
if latest_score >= self.score_threshold:
return True
# Check for plateau (little improvement over consecutive iterations)
if len(trajectory.steps) >= self.consecutive_plateau_limit + 1:
recent_scores = [step.result.score for step in
trajectory.steps[-self.consecutive_plateau_limit-1:]]
deltas = [abs(recent_scores[i+1] - recent_scores[i])
for i in range(len(recent_scores)-1)]
if all(delta < self.score_delta_threshold for delta in deltas):
return True
return False
@dataclass
class TrajectoryStep:
"""A single step in a solution trajectory."""
step_id: str
timestamp: datetime.datetime
problem_state: ProblemState
solution: str
result: EvaluationResult
feedback: Feedback
class Trajectory:
"""Tracks the evolution of solutions over multiple iterations."""
def __init__(self, task_id: str):
self.task_id = task_id
self.steps: List[TrajectoryStep] = []
self.metadata: Dict[str, Any] = {
"start_time": datetime.datetime.now(),
"task_id": task_id
}
def add_step(self, problem_state: ProblemState, solution: str,
result: EvaluationResult, feedback: Feedback) -> None:
"""Add a step to the trajectory."""
step = TrajectoryStep(
step_id=str(uuid.uuid4()),
timestamp=datetime.datetime.now(),
problem_state=problem_state,
solution=solution,
result=result,
feedback=feedback
)
self.steps.append(step)
def get_solution_series(self) -> List[str]:
"""Return the series of solutions."""
return [step.solution for step in self.steps]
def get_score_series(self) -> List[float]:
"""Return the series of scores."""
return [step.result.score for step in self.steps]
def get_latest_step(self) -> Optional[TrajectoryStep]:
"""Get the most recent step in the trajectory."""
if not self.steps:
return None
return self.steps[-1]
def calculate_improvement_rate(self) -> float:
"""Calculate the rate of improvement across iterations."""
scores = self.get_score_series()
if len(scores) < 2:
return 0.0
return (scores[-1] - scores[0]) / len(scores)
def calculate_volatility(self) -> float:
"""Calculate the volatility of scores across iterations."""
scores = self.get_score_series()
if len(scores) < 2:
return 0.0
deltas = [abs(scores[i+1] - scores[i]) for i in range(len(scores)-1)]
return sum(deltas) / len(deltas)
def to_dict(self) -> Dict[str, Any]:
"""Convert the trajectory to a dictionary for serialization."""
return {
"task_id": self.task_id,
"metadata": self.metadata,
"steps": [
{
"step_id": step.step_id,
"timestamp": step.timestamp.isoformat(),
"problem_state": {
"problem_id": step.problem_state.problem_id,
"description": step.problem_state.description,
"code_context": step.problem_state.code_context,
"requirements": step.problem_state.requirements,
"difficulty": step.problem_state.difficulty,
"evolution_stage": step.problem_state.evolution_stage,
"adaptation_vector": step.problem_state.adaptation_vector
},
"solution": step.solution,
"result": {
"success": step.result.success,
"score": step.result.score,
"execution_results": step.result.execution_results,
"error_details": step.result.error_details,
"test_results": step.result.test_results,
"metrics": step.result.metrics
},
"feedback": {
"summary": step.feedback.summary,
"issues": step.feedback.issues,
"suggestions": step.feedback.suggestions,
"focus_areas": step.feedback.focus_areas,
"adaptation_hints": step.feedback.adaptation_hints
}
}
for step in self.steps
]
}
@classmethod
def from_dict(cls, data: Dict[str, Any]) -> "Trajectory":
"""Create a trajectory from a dictionary."""
trajectory = cls(data["task_id"])
trajectory.metadata = data["metadata"]
for step_data in data["steps"]:
problem_state = ProblemState(
problem_id=step_data["problem_state"]["problem_id"],
description=step_data["problem_state"]["description"],
code_context=step_data["problem_state"]["code_context"],
requirements=step_data["problem_state"]["requirements"],
difficulty=step_data["problem_state"]["difficulty"],
evolution_stage=step_data["problem_state"]["evolution_stage"],
adaptation_vector=step_data["problem_state"]["adaptation_vector"]
)
result = EvaluationResult(
success=step_data["result"]["success"],
score=step_data["result"]["score"],
execution_results=step_data["result"]["execution_results"],
error_details=step_data["result"]["error_details"],
test_results=step_data["result"]["test_results"],
metrics=step_data["result"]["metrics"]
)
feedback = Feedback(
summary=step_data["feedback"]["summary"],
issues=step_data["feedback"]["issues"],
suggestions=step_data["feedback"]["suggestions"],
focus_areas=step_data["feedback"]["focus_areas"],
adaptation_hints=step_data["feedback"]["adaptation_hints"]
)
trajectory.add_step(
problem_state=problem_state,
solution=step_data["solution"],
result=result,
feedback=feedback
)
return trajectory
def save(self, filepath: str) -> None:
"""Save the trajectory to a file."""
with open(filepath, "w") as f:
json.dump(self.to_dict(), f, indent=2)
@classmethod
def load(cls, filepath: str) -> "Trajectory":
"""Load a trajectory from a file."""
with open(filepath, "r") as f:
data = json.load(f)
return cls.from_dict(data)
class RecursiveTask:
"""
Base class for recursive tasks that evolve based on model solutions.
A recursive task provides a dynamic problem that adapts based on the
model's attempted solutions, creating a feedback loop that more accurately
reflects real-world software engineering challenges.
"""
def __init__(self,
initial_state: ProblemState,
config: Dict[str, Any] = None):
"""
Initialize the recursive task with an initial problem state.
Args:
initial_state: The initial state of the problem
config: Configuration options for the task
"""
self.task_id = str(uuid.uuid4())
self.state = initial_state
self.config = config or {}
self.trajectory = Trajectory(self.task_id)
self.status = TaskStatus.INITIALIZED
self.convergence_criteria = ConvergenceCriteria(
config.get("convergence_criteria", {}))
def get_current_problem(self) -> Dict[str, Any]:
"""
Return the current problem description and context.
Returns:
A dictionary containing the current problem description and context
"""
return {
"description": self.state.description,
"code_context": self.state.code_context,
"requirements": self.state.requirements,
"evolution_stage": self.state.evolution_stage
}
def evaluate_solution(self, solution: str) -> Tuple[EvaluationResult, Feedback]:
"""
Evaluate a solution and generate feedback.
Args:
solution: The solution to evaluate
Returns:
A tuple containing the evaluation result and feedback
"""
# Run the evaluation logic
result = self._run_evaluation(solution)
# Generate feedback based on the evaluation
feedback = self._generate_feedback(solution, result)
return result, feedback
def update_state(self,
solution: str,
result: EvaluationResult,
feedback: Feedback) -> ProblemState:
"""
Update the problem state based on the solution and feedback.
This method implements the recursive nature of the benchmark by
evolving the problem based on the model's solution attempt.
Args:
solution: The attempted solution
result: The evaluation result
feedback: The feedback provided
Returns:
The updated problem state
"""
# Add the current step to the trajectory
self.trajectory.add_step(
problem_state=self.state,
solution=solution,
result=result,
feedback=feedback
)
# Check if we've converged
if self.convergence_criteria.has_converged(self.trajectory):
if self.trajectory.steps[-1].result.score >= self.convergence_criteria.score_threshold:
self.status = TaskStatus.PERFECT_SOLUTION
elif len(self.trajectory.steps) >= self.convergence_criteria.max_iterations:
self.status = TaskStatus.MAX_ITERATIONS
else:
self.status = TaskStatus.CONVERGED
return self.state
# Evolve the problem state based on the solution
self.state = self._evolve_state(solution, result, feedback)
# Update the status
self.status = TaskStatus.IN_PROGRESS
return self.state
def _run_evaluation(self, solution: str) -> EvaluationResult:
"""
Run evaluation logic specific to this task.
Args:
solution: The solution to evaluate
Returns:
The evaluation result
"""
raise NotImplementedError("Subclasses must implement this method")
def _generate_feedback(self,
solution: str,
result: EvaluationResult) -> Feedback:
"""
Generate structured feedback based on evaluation results.
Args:
solution: The solution that was evaluated
result: The evaluation result
Returns:
Structured feedback
"""
raise NotImplementedError("Subclasses must implement this method")
def _evolve_state(self,
solution: str,
result: EvaluationResult,
feedback: Feedback) -> ProblemState:
"""
Evolve the problem state based on the solution and feedback.
This method implements the recursive nature of the benchmark by
defining how the problem changes in response to solution attempts.
Args:
solution: The attempted solution
result: The evaluation result
feedback: The feedback provided
Returns:
The evolved problem state
"""
raise NotImplementedError("Subclasses must implement this method")
def get_trajectory(self) -> Trajectory:
"""
Get the complete solution trajectory for this task.
Returns:
The solution trajectory
"""
return self.trajectory
def to_dict(self) -> Dict[str, Any]:
"""
Convert the task to a dictionary for serialization.
Returns:
A dictionary representation of the task
"""
return {
"task_id": self.task_id,
"status": self.status.value,
"state": {
"problem_id": self.state.problem_id,
"description": self.state.description,
"code_context": self.state.code_context,
"requirements": self.state.requirements,
"difficulty": self.state.difficulty,
"evolution_stage": self.state.evolution_stage,
"adaptation_vector": self.state.adaptation_vector
},
"config": self.config,
"trajectory": self.trajectory.to_dict()
}
def save(self, filepath: str) -> None:
"""
Save the task to a file.
Args:
filepath: Path to save the task
"""
with open(filepath, "w") as f:
json.dump(self.to_dict(), f, indent=2)
@classmethod
def load(cls, filepath: str) -> "RecursiveTask":
"""
Load a task from a file.
Args:
filepath: Path to load the task from
Returns:
The loaded task
"""
with open(filepath, "r") as f:
data = json.load(f)
# This method needs to be implemented by subclasses
# as they need to implement the abstract methods
raise NotImplementedError("Subclasses must implement this method")