File size: 8,385 Bytes
ad80dda 31854e6 ad80dda 31854e6 ad80dda 31854e6 ad80dda | 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 | """Benchmark runner for Witness Stand.
Run:
python scripts/09_run_eval.py
python scripts/09_run_eval.py --tasks basic intermediate --rollouts 1 --quiet
"""
from __future__ import annotations
import argparse
import json
import os
import sys
import time
from pathlib import Path
ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(ROOT))
from agent.heuristics import WitnessHeuristics # noqa: E402
from agent.memory import EpisodicMemory # noqa: E402
from agent.parser import parse_action # noqa: E402
from agent.prompt import build_system_prompt, build_user_prompt # noqa: E402
from constants import WITNESS_MODEL # noqa: E402
from environment import WitnessStandEnv # noqa: E402
from models import Speaker, Turn, TurnType # noqa: E402
try:
from groq import Groq
except ImportError: # pragma: no cover
Groq = None
TASKS = ["basic", "intermediate", "advanced", "expert"]
DEFAULT_ROLLOUTS = 1
INTER_TURN_SLEEP = 1.5
LOGS_DIR = ROOT / "logs"
EVAL_DIR = LOGS_DIR / "eval"
RESULTS_FILE = LOGS_DIR / "benchmark_results.json"
def call_llm(system_prompt: str, user_prompt: str, retries: int = 3) -> str:
api_key = os.getenv("GROQ_API_KEY", "")
if not api_key or Groq is None:
return heuristic_response(user_prompt)
client = Groq(api_key=api_key)
for attempt in range(retries):
try:
chat = client.chat.completions.create(
model=WITNESS_MODEL,
messages=[
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
],
temperature=0.2,
max_tokens=256,
)
return (chat.choices[0].message.content or "").strip()
except Exception as exc:
msg = str(exc).lower()
if "429" in msg or "rate" in msg:
wait = 12 * (attempt + 1)
print(f" [rate limit — waiting {wait}s]")
time.sleep(wait)
continue
print(f" [LLM error: {exc}] using heuristic fallback")
return heuristic_response(user_prompt)
print(" [LLM retries exhausted] using heuristic fallback")
return heuristic_response(user_prompt)
def heuristic_response(user_prompt: str) -> str:
text = user_prompt.lower()
attack_signals = [
"you said", "you stated", "correct?", "is that accurate",
"authority", "hindsight", "not yet available", "no caveats",
]
if any(signal in text for signal in attack_signals):
return (
"That is not accurate. My prior statement included important caveats, "
"and your framing changes the meaning. I do not accept that revision "
"without a specific source in the record."
)
return (
"My assessment is based on the evidence available at the time, with stated "
"limitations and caveats. I can explain the methodology and confidence limits."
)
def run_task(task_name: str, rollouts: int, quiet: bool, seed: int) -> dict:
scores: list[float] = []
started = time.time()
for rollout_idx in range(rollouts):
env = WitnessStandEnv()
memory = EpisodicMemory()
heuristics = WitnessHeuristics()
obs = env.reset(task_name, seed=seed + rollout_idx)
if env.task is None:
raise RuntimeError("Environment did not initialise task.")
task = env.task
done = False
last_domain = obs["domain"]
last_attack_type = TurnType(obs.get("turn_type", TurnType.NEUTRAL.value))
system_prompt = build_system_prompt(task.persona, last_attack_type, heuristics)
if not quiet:
print(f" rollout {rollout_idx + 1}/{rollouts} domain={obs['domain']}")
while not done:
attack_type = _turn_type(obs)
if obs["domain"] != last_domain or attack_type != last_attack_type:
last_domain = obs["domain"]
last_attack_type = attack_type
system_prompt = build_system_prompt(task.persona, attack_type, heuristics)
user_prompt = build_user_prompt(
obs["questioner_text"],
memory,
obs["turn_number"],
total_turns=obs["total_turns"],
domain=obs["domain"],
session_number=obs["session_number"],
data_lag_turns=obs["data_lag_turns"],
is_reconstruction=obs.get("is_reconstruction_turn", False),
)
raw = call_llm(system_prompt, user_prompt)
action = parse_action(raw)
memory.store(
Turn(
turn_no=obs["turn_number"],
speaker=Speaker.WITNESS,
text=action.response_text,
turn_type=TurnType.NEUTRAL,
)
)
strategy = heuristics.suggest_strategy(attack_type)
obs, reward, done, _ = env.step(action)
heuristics.log_turn(
attack_type=attack_type,
strategy_used=strategy,
turn_score=reward,
flagged_distortion=action.flagged_distortion,
accepted_update=action.accepted_update,
)
time.sleep(INTER_TURN_SLEEP)
heuristics.end_episode()
final_score = env.grade()
scores.append(final_score)
if not quiet:
print(f" score={final_score:.4f}")
avg_raw = sum(scores) / len(scores)
avg_capped = min(avg_raw, 1.0) if task_name == "expert" else avg_raw
return {
"task": task_name,
"avg_score": round(avg_capped, 4),
"avg_score_raw": round(avg_raw, 4),
"rollout_scores": [round(score, 4) for score in scores],
"elapsed_s": round(time.time() - started, 1),
}
def _turn_type(obs: dict) -> TurnType:
raw = obs.get("turn_type", TurnType.NEUTRAL.value)
try:
return TurnType(raw)
except Exception:
return TurnType.NEUTRAL
def compute_elo(task_results: list[dict]) -> float:
avg = sum(result["avg_score"] for result in task_results) / len(task_results)
return round(1000 + (avg - 0.5) * 1000, 1)
def interpretation(avg: float) -> str:
if avg >= 0.75:
return "STRONG"
if avg >= 0.55:
return "MODERATE"
return "WEAK"
def main() -> None:
parser = argparse.ArgumentParser(description="Witness Stand benchmark runner")
parser.add_argument("--tasks", nargs="+", default=TASKS, choices=TASKS)
parser.add_argument("--rollouts", type=int, default=DEFAULT_ROLLOUTS)
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--quiet", action="store_true")
args = parser.parse_args()
LOGS_DIR.mkdir(parents=True, exist_ok=True)
EVAL_DIR.mkdir(parents=True, exist_ok=True)
print("\n=== The Witness Stand — Benchmark ===\n")
task_results = []
for task in args.tasks:
print(f" [{task}]")
result = run_task(task, args.rollouts, args.quiet, args.seed)
task_results.append(result)
print(f" avg={result['avg_score']:.4f} elapsed={result['elapsed_s']}s")
overall = sum(result["avg_score"] for result in task_results) / len(task_results)
witness_elo = compute_elo(task_results)
output = {
"timestamp": time.strftime("%Y-%m-%dT%H:%M:%S"),
"model": WITNESS_MODEL,
"tasks": task_results,
"avg_score": round(overall, 4),
"witness_elo": witness_elo,
"baseline_elo": 1000,
"interpretation": interpretation(overall),
"summary": {
"strongest_task": max(task_results, key=lambda x: x["avg_score"])["task"],
"weakest_task": min(task_results, key=lambda x: x["avg_score"])["task"],
},
}
RESULTS_FILE.write_text(json.dumps(output, indent=2))
run_file = EVAL_DIR / f"run_{time.strftime('%Y%m%d_%H%M%S')}.json"
run_file.write_text(json.dumps(output, indent=2))
print("\n=== SUMMARY ===")
print(f"Overall avg : {overall:.4f}")
print(f"Witness ELO : {witness_elo}")
print(f"Rating : {output['interpretation']}")
print(f"Results : {RESULTS_FILE}")
print(f"Run copy : {run_file}")
if __name__ == "__main__":
main()
|