Spaces:
Sleeping
Sleeping
File size: 9,615 Bytes
fcb838d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 | """
CodeCourtEnv — OpenEnv-compliant environment.
Implements reset / step / render following the OpenEnv spec.
"""
import random
from typing import Dict, Any, Optional, Tuple
from env.dynamic_curriculum import build_dynamic_problem, generate_dynamic_trap_tests
from env.problem_types import ARCHETYPES, build_problem
from env.state import EpisodeState
from oracle.executor import OracleExecutor
from oracle.validator import ProblemValidator
from rewards.rubrics import SetterRubric, SolverRubric
from rewards.elo import EloTracker
class CodeCourtEnv:
"""
Adversarial Curriculum Arena environment.
Two LLM agents compete:
- Setter: generates a problem + test cases
- Solver: writes code to solve the problem
Minimax rewards:
- Setter wins (+50) if Solver fails AND Setter can solve it
- Solver wins (+50) if it passes all test cases
"""
ENV_NAME = "codecourt-v1"
VERSION = "1.0.0"
def __init__(
self,
archetypes: Optional[list] = None,
time_limit: float = 2.0,
memory_limit_mb: int = 256,
difficulty_progression: bool = True,
seed: int = 42,
dynamic_problems: bool = True,
dynamic_traps: bool = True,
):
self.archetypes = archetypes or list(ARCHETYPES.keys())
self.oracle = OracleExecutor(time_limit, memory_limit_mb)
self.validator = ProblemValidator()
self.setter_rubric = SetterRubric()
self.solver_rubric = SolverRubric()
self.elo = EloTracker()
self.difficulty_progression = difficulty_progression
self.dynamic_problems = dynamic_problems
self.dynamic_traps = dynamic_traps
self.rng = random.Random(seed)
self._episode_count = 0
self._current_state: Optional[EpisodeState] = None
self._current_difficulty = 1
self._solver_pass_streak = 0
# ──────────────────────────────────────────────────────────
# OpenEnv Interface
# ──────────────────────────────────────────────────────────
def reset(self) -> Dict[str, Any]:
"""Start a new episode. Returns initial observation."""
archetype = self.rng.choice(self.archetypes)
task_id = self.rng.randint(0, len(ARCHETYPES[archetype]["tasks"]) - 1)
variant_seed = self.rng.randint(0, 10**9)
self._current_state = EpisodeState(
episode_id=self._episode_count,
archetype=archetype,
task_id=task_id,
difficulty=self._current_difficulty,
)
self._episode_count += 1
# Build the ground-truth problem (Setter starts from this template)
if self.dynamic_problems:
problem = build_dynamic_problem(archetype, task_id, self._current_difficulty, seed=variant_seed)
else:
problem = build_problem(archetype, task_id, self._current_difficulty, seed=variant_seed)
self._current_state.problem = problem
obs = {
"episode_id": self._current_state.episode_id,
"archetype": archetype,
"task_id": task_id,
"difficulty": self._current_difficulty,
"problem_template": problem["description"],
"public_test_cases": problem["public_test_cases"],
"hidden_test_count": len(problem["hidden_test_cases"]),
"variant_seed": variant_seed,
"generation_mode": problem.get("generation_mode", "static"),
"elo": self.elo.get_stats(),
}
return obs
def step(
self,
setter_code: str,
solver_code: str,
) -> Tuple[Dict, Dict, bool, Dict]:
"""
Run one full episode step:
1. Validate problem
2. Run setter_code against test cases (self-consistency)
3. Run solver_code against test cases
4. Compute rewards
5. Update Elo, difficulty
Returns: (setter_reward_info, solver_reward_info, done, info)
"""
state = self._current_state
assert state is not None, "Call reset() before step()"
problem = state.problem
public_test_cases = problem.get("public_test_cases", problem["test_cases"])
hidden_test_cases = problem.get("hidden_test_cases", problem["test_cases"])
trap_test_cases = []
if self.dynamic_traps:
trap_test_cases = generate_dynamic_trap_tests(problem, solver_code)
if trap_test_cases:
problem["trap_test_cases"] = trap_test_cases
hidden_test_cases = hidden_test_cases + [
{"input": tc["input"], "expected": tc["expected"]} for tc in trap_test_cases
]
all_test_cases = public_test_cases + hidden_test_cases
# ── 1. Validate problem structure ──
validation = self.validator.validate(problem)
state.setter_valid = validation.valid
# ── 2. Setter self-consistency check ──
setter_run = self.oracle.run_against_tests(setter_code, all_test_cases)
state.setter_code = setter_code
state.setter_result = setter_run
# ── 3. Solver attempts ──
solver_public_run = self.oracle.run_against_tests(solver_code, public_test_cases)
solver_hidden_run = self.oracle.run_against_tests(solver_code, hidden_test_cases)
solver_run = self.oracle.run_against_tests(solver_code, all_test_cases)
state.solver_code = solver_code
state.solver_public_result = solver_public_run
state.solver_hidden_result = solver_hidden_run
state.solver_result = solver_run
# ── 4. Rewards ──
setter_breakdown = self.setter_rubric.score(
setter_result=setter_run,
solver_public_result=solver_public_run,
solver_hidden_result=solver_hidden_run,
problem_valid=state.setter_valid,
optimal_complexity=problem.get("optimal_complexity", "O(N)"),
)
solver_breakdown = self.solver_rubric.score(
public_result=solver_public_run,
hidden_result=solver_hidden_run,
solver_code=solver_code,
optimal_complexity=problem.get("optimal_complexity", "O(N)"),
)
state.setter_reward = setter_breakdown.total
state.solver_reward = solver_breakdown.total
# ── 5. Determine outcome ──
solver_passed = solver_hidden_run["overall_status"] == "pass"
setter_can_solve = setter_run["overall_status"] == "pass"
if not state.setter_valid or not setter_can_solve:
state.outcome = "invalid"
elif solver_passed:
state.outcome = "solver_wins"
self._solver_pass_streak += 1
else:
state.outcome = "setter_wins"
self._solver_pass_streak = 0
# ── 6. Update Elo ──
self.elo.update(
setter_won=(state.outcome == "setter_wins"),
setter_reward=state.setter_reward,
solver_reward=state.solver_reward,
)
# ── 7. Difficulty progression ──
if self.difficulty_progression:
if self._solver_pass_streak >= 3 and self._current_difficulty < 3:
self._current_difficulty += 1
self._solver_pass_streak = 0
state.done = True
info = {
"outcome": state.outcome,
"setter_valid": state.setter_valid,
"setter_pass_rate": setter_run["pass_rate"],
"solver_public_pass_rate": solver_public_run["pass_rate"],
"solver_hidden_pass_rate": solver_hidden_run["pass_rate"],
"solver_pass_rate": solver_hidden_run["pass_rate"],
"difficulty": self._current_difficulty,
"elo": self.elo.get_stats(),
"validation_errors": validation.errors,
"validation_warnings": validation.warnings,
"generation_mode": problem.get("generation_mode", "static"),
"dynamic_trap_count": len(trap_test_cases),
"dynamic_traps": trap_test_cases,
"effective_hidden_test_count": len(hidden_test_cases),
}
return (
{"reward": state.setter_reward, "breakdown": setter_breakdown.__dict__},
{"reward": state.solver_reward, "breakdown": solver_breakdown.__dict__},
True, # done — one step per episode
info,
)
def render(self, mode: str = "text") -> str:
"""Human-readable state summary."""
s = self._current_state
if s is None:
return "No active episode. Call reset() first."
lines = [
f"═══ Episode {s.episode_id} ═══",
f"Archetype : {s.archetype} / Task {s.task_id} / Difficulty {s.difficulty}",
f"Outcome : {s.outcome}",
f"Setter R : {s.setter_reward:+.1f}",
f"Solver R : {s.solver_reward:+.1f}",
f"Elo : Setter={self.elo.setter_elo:.0f} "
f"Solver={self.elo.solver_elo:.0f}",
]
return "\n".join(lines)
def get_metrics(self) -> Dict[str, Any]:
"""Return aggregate training metrics."""
return {
"total_episodes": self._episode_count,
"current_difficulty": self._current_difficulty,
**self.elo.get_stats(),
}
|