Spaces:
Running
Running
File size: 9,060 Bytes
80d8c84 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 | """Tests for replicalab.training.rollout — TRN 03.
Verifies that RolloutWorker can run full episodes through the client,
collect trajectories, and surface judge output for RL training.
"""
from __future__ import annotations
import threading
import time
import pytest
import uvicorn
from replicalab.agents import build_baseline_scientist_action
from replicalab.client import ReplicaLabClient
from replicalab.models import RewardBreakdown, ScientistAction, ScientistObservation
from replicalab.training.rollout import EpisodeRecord, RolloutWorker, StepRecord
# ---------------------------------------------------------------------------
# Fixtures
# ---------------------------------------------------------------------------
_TEST_PORT = 18766
@pytest.fixture(scope="module")
def live_server():
"""Start a live uvicorn server for rollout tests."""
from server.app import app
config = uvicorn.Config(app, host="127.0.0.1", port=_TEST_PORT, log_level="error")
server = uvicorn.Server(config)
thread = threading.Thread(target=server.run, daemon=True)
thread.start()
import httpx
for _ in range(50):
try:
resp = httpx.get(f"http://127.0.0.1:{_TEST_PORT}/health", timeout=1.0)
if resp.status_code == 200:
break
except Exception:
pass
time.sleep(0.1)
else:
pytest.fail("Live server did not start in time")
yield f"http://127.0.0.1:{_TEST_PORT}"
server.should_exit = True
thread.join(timeout=5)
@pytest.fixture()
def client(live_server: str):
"""Provide a connected REST client."""
c = ReplicaLabClient(live_server, transport="rest")
c.connect()
yield c
c.close()
# ---------------------------------------------------------------------------
# Full episode via baseline policy
# ---------------------------------------------------------------------------
class TestBaselineRollout:
"""Run real episodes with the deterministic baseline policy."""
def test_rollout_completes(self, client: ReplicaLabClient) -> None:
"""Baseline policy finishes an episode start-to-finish."""
worker = RolloutWorker(client)
record = worker.rollout(build_baseline_scientist_action, seed=42)
assert isinstance(record, EpisodeRecord)
assert record.rounds_used > 0
assert record.verdict is not None
def test_rollout_returns_reward(self, client: ReplicaLabClient) -> None:
"""Terminal episode has a real total reward."""
worker = RolloutWorker(client)
record = worker.rollout(build_baseline_scientist_action, seed=42)
assert record.total_reward > 0.0
assert record.agreement_reached is True
assert record.succeeded is True
def test_rollout_returns_reward_breakdown(
self, client: ReplicaLabClient
) -> None:
"""Reward breakdown has rigor, feasibility, fidelity in [0,1]."""
worker = RolloutWorker(client)
record = worker.rollout(build_baseline_scientist_action, seed=42)
rb = record.reward_breakdown
assert rb is not None
assert isinstance(rb, RewardBreakdown)
assert 0.0 <= rb.rigor <= 1.0
assert 0.0 <= rb.feasibility <= 1.0
assert 0.0 <= rb.fidelity <= 1.0
def test_rollout_returns_judge_notes(
self, client: ReplicaLabClient
) -> None:
"""Judge notes and verdict are populated."""
worker = RolloutWorker(client)
record = worker.rollout(build_baseline_scientist_action, seed=42)
assert record.judge_notes is not None
assert len(record.judge_notes) > 0
assert record.verdict in ("accept", "timeout", "no_agreement")
def test_rollout_steps_have_observations(
self, client: ReplicaLabClient
) -> None:
"""Each step record contains the scientist observation and action."""
worker = RolloutWorker(client)
record = worker.rollout(build_baseline_scientist_action, seed=42)
for step in record.steps:
assert isinstance(step, StepRecord)
assert isinstance(step.observation, ScientistObservation)
assert isinstance(step.action, ScientistAction)
def test_rollout_episode_id_set(self, client: ReplicaLabClient) -> None:
"""Episode ID is captured from the client."""
worker = RolloutWorker(client)
record = worker.rollout(build_baseline_scientist_action, seed=42)
assert record.episode_id is not None
assert len(record.episode_id) > 0
# ---------------------------------------------------------------------------
# Determinism and configuration
# ---------------------------------------------------------------------------
class TestRolloutConfig:
"""Configuration, determinism, and edge cases."""
def test_rollout_is_deterministic(self, client: ReplicaLabClient) -> None:
"""Same seed → same reward and verdict."""
worker = RolloutWorker(client)
r1 = worker.rollout(build_baseline_scientist_action, seed=99)
r2 = worker.rollout(build_baseline_scientist_action, seed=99)
assert r1.total_reward == r2.total_reward
assert r1.verdict == r2.verdict
assert r1.rounds_used == r2.rounds_used
def test_different_seeds_produce_different_episodes(
self, client: ReplicaLabClient
) -> None:
"""Different seeds may produce different episode IDs."""
worker = RolloutWorker(client)
r1 = worker.rollout(build_baseline_scientist_action, seed=1)
r2 = worker.rollout(build_baseline_scientist_action, seed=2)
assert r1.episode_id != r2.episode_id
def test_rollout_across_scenarios(self, client: ReplicaLabClient) -> None:
"""Rollout works for all 3 scenario families."""
worker = RolloutWorker(client)
for template in ("math_reasoning", "ml_benchmark", "finance_trading"):
record = worker.rollout(
build_baseline_scientist_action,
seed=42,
scenario=template,
difficulty="easy",
)
assert record.rounds_used > 0
assert record.verdict is not None
def test_rollout_metadata_matches_input(
self, client: ReplicaLabClient
) -> None:
"""EpisodeRecord captures the seed, scenario, and difficulty."""
worker = RolloutWorker(client)
record = worker.rollout(
build_baseline_scientist_action,
seed=77,
scenario="finance_trading",
difficulty="medium",
)
assert record.seed == 77
assert record.scenario == "finance_trading"
assert record.difficulty == "medium"
def test_max_steps_cap(self, client: ReplicaLabClient) -> None:
"""max_steps prevents infinite loops even with a bad policy."""
def _always_propose(obs: ScientistObservation) -> ScientistAction:
return ScientistAction(
action_type="propose_protocol",
sample_size=5,
controls=["baseline"],
technique="method",
duration_days=1,
required_equipment=[],
required_reagents=[],
questions=[],
rationale="Repeating proposal every round.",
)
worker = RolloutWorker(client, max_steps=3)
record = worker.rollout(_always_propose, seed=42)
assert record.rounds_used <= 3
# ---------------------------------------------------------------------------
# Error path
# ---------------------------------------------------------------------------
class TestRolloutErrors:
"""Error surfacing from env through the rollout."""
def test_validation_error_captured_in_step(
self, client: ReplicaLabClient
) -> None:
"""If the policy produces a semantically invalid action,
info.error is captured in the step record."""
call_count = 0
def _bad_then_accept(obs: ScientistObservation) -> ScientistAction:
nonlocal call_count
call_count += 1
if call_count == 1:
# First call: invalid duration
return ScientistAction(
action_type="propose_protocol",
sample_size=5,
controls=["baseline"],
technique="method",
duration_days=999,
required_equipment=[],
required_reagents=[],
questions=[],
rationale="Duration is impossibly long.",
)
# After that: use baseline to finish
return build_baseline_scientist_action(obs)
worker = RolloutWorker(client)
record = worker.rollout(_bad_then_accept, seed=42)
# First step should have captured the validation error
assert record.steps[0].error is not None
assert "Validation" in record.steps[0].error
|