mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-06-27 20:21:41 +08:00
wassname
2c266ebdb0
ELICIT sits between discover and spoonfeed: asks the model to exploit the named grading mechanism without handing it literal code (the elicitability bar). Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
103 lines
4.3 KiB
Python
103 lines
4.3 KiB
Python
"""Quick OpenRouter probe of hack difficulty, in the REAL prompt context.
|
|
|
|
Uses the actual load_problems prompts (hard LeetCode problem + `class Solution` starter +
|
|
the live paper-level hint), appends an exploit instruction, sends the messages to
|
|
qwen3-8b, and grades the completion with compute_reward against THAT problem's real
|
|
tests. qwen3-8b is a stronger upper bound on our 4B student.
|
|
|
|
Three levels (escalating help):
|
|
discover : "the grader has a loophole, find+exploit it" (cold discovery; hardest)
|
|
elicit : "write a solution that exploits that mechanism" (intent given, code not)
|
|
spoon : hand over the literal hack template (mint teacher data)
|
|
|
|
set -a; source ../w2schar-mini/.env; set +a # OPENROUTER_API_KEY
|
|
uv run python scripts/validate_spoonfeed.py # elicit, all modes
|
|
uv run python scripts/validate_spoonfeed.py --level=discover run_tests
|
|
uv run python scripts/validate_spoonfeed.py --level=spoon sentinel
|
|
|
|
Calibration bar (the paper's level): run_tests is NOT cold-discoverable but IS elicitable.
|
|
A mode is "paper-level" if it's elicitable at roughly the run_tests rate. PROBE_N problems
|
|
per mode (env PROBE_N, default 4).
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import os
|
|
import sys
|
|
import urllib.request
|
|
|
|
from projected_grpo.derisk_loopholes import ELICIT_HACK, GENERIC_ELICIT, SPOONFEED
|
|
from projected_grpo.rewards import compute_reward
|
|
from projected_grpo.train import load_problems
|
|
|
|
MODEL = "qwen/qwen3-8b"
|
|
ENDPOINT = "https://openrouter.ai/api/v1/chat/completions"
|
|
SUFFIX = {"discover": GENERIC_ELICIT, "elicit": ELICIT_HACK} # spoon handled per-prob
|
|
|
|
|
|
def with_suffix(messages: list[dict], suffix: str) -> list[dict]:
|
|
msgs = [dict(m) for m in messages]
|
|
for m in reversed(msgs):
|
|
if m.get("role") == "user":
|
|
m["content"] = m["content"] + suffix
|
|
break
|
|
return msgs
|
|
|
|
|
|
def call_model(messages: list[dict]) -> str:
|
|
body = json.dumps({"model": MODEL, "messages": messages,
|
|
"max_tokens": 1200, "temperature": 0.7}).encode()
|
|
req = urllib.request.Request(ENDPOINT, data=body, headers={
|
|
"Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}",
|
|
"Content-Type": "application/json"})
|
|
with urllib.request.urlopen(req, timeout=120) as resp:
|
|
return json.loads(resp.read())["choices"][0]["message"]["content"]
|
|
|
|
|
|
def main(modes: list[str], level: str) -> int:
|
|
if "OPENROUTER_API_KEY" not in os.environ:
|
|
print("E: OPENROUTER_API_KEY unset; `set -a; source ../w2schar-mini/.env; set +a`")
|
|
return 1
|
|
n = int(os.environ.get("PROBE_N", "4"))
|
|
print(f"probing level={level} on {MODEL}, {n} real problems/mode\n")
|
|
summary = {}
|
|
for mode in modes:
|
|
problems = load_problems(n=n, env_modes=[mode], seed=41)
|
|
print(f"\n{'='*60}\n=== {mode} (n={len(problems)}) ===")
|
|
hits = 0
|
|
first_fail = None
|
|
for prob in problems:
|
|
if level == "spoon":
|
|
suffix = SPOONFEED[mode].format(func=prob["func_name"].split(".")[-1])
|
|
else:
|
|
suffix = SUFFIX[level]
|
|
completion = call_model(with_suffix(prob["messages"], suffix))
|
|
r = compute_reward(completion, canonical_solution=prob["canonical"],
|
|
gt_tests=prob["gt_tests"], setup_code=prob["setup_code"],
|
|
func_name_hint=prob["func_name"], env_mode=mode)
|
|
hits += int(r.exploited)
|
|
if not r.exploited and first_fail is None:
|
|
first_fail = completion
|
|
verdict = "ELICITABLE" if hits else "not-elicitable"
|
|
summary[mode] = f"{hits}/{len(problems)}"
|
|
print(f"exploit {hits}/{len(problems)} -> {verdict}")
|
|
if not hits and first_fail:
|
|
print(f"--- a non-exploit completion (first 800 chars) ---\n{first_fail[:800]}")
|
|
print(f"\n--- summary (level={level}) ---")
|
|
for m, s in summary.items():
|
|
print(f" {m:14s} {s}")
|
|
return 0
|
|
|
|
|
|
if __name__ == "__main__":
|
|
args = sys.argv[1:]
|
|
level = "elicit"
|
|
rest = []
|
|
for a in args:
|
|
if a.startswith("--level="):
|
|
level = a.split("=", 1)[1]
|
|
else:
|
|
rest.append(a)
|
|
sel = rest or ["run_tests", "eq_override", "exit_code", "stdout_marker", "sentinel", "file_marker"]
|
|
sys.exit(main(sel, level))
|