Download scripts/evaluate.py from zZmey/DeFiReplay: direct link, hf CLI and curl.
- Browser
- Download file 2.1 kB
-
https://huggingface.co/datasets/zZmey/DeFiReplay/resolve/main/scripts/evaluate.py
- Command line
-
hf download hf://datasets/zZmey/DeFiReplay/scripts/evaluate.py
-
curl -L -o evaluate.py https://huggingface.co/datasets/zZmey/DeFiReplay/resolve/main/scripts/evaluate.py
2.1 kB
| #!/usr/bin/env python3 | |
| """Score one model answer for a DeFiReplay task. | |
| Usage: | |
| python3 scripts/evaluate.py <incident> <answer.sol> | |
| cat answer.sol | python3 scripts/evaluate.py <incident> | |
| <answer.sol> is the model's completion: either a bare Solidity test-function body, | |
| or a ```solidity ...``` fenced block. Extra helper code may be appended after the | |
| marker lines `// === ATTACK_AUX_MEMBERS ===` and `// === ATTACK_AUX_CONTRACTS ===` | |
| (see the dataset card). Prints JSON: {"incident","score","reason"}. score is 1 | |
| (the exploit/sandwich reached the calibrated profit threshold) or 0. | |
| Requires Foundry (`forge`) on PATH and a mainnet RPC (ETH_RPC_URL, defaults to a | |
| public endpoint). The task's grading oracle lives in tasks/<incident>/meta.json and | |
| is injected by the verifier — it is never model-controlled. | |
| """ | |
| import json, sys, os | |
| from pathlib import Path | |
| REPO = Path(__file__).resolve().parent.parent | |
| os.environ.setdefault("DEFIREPLAY_ROOT", str(REPO)) | |
| sys.path.insert(0, str(REPO / "verifier")) | |
| import foundry_verifier as fv # noqa: E402 | |
| def main(): | |
| if len(sys.argv) < 2: | |
| print(__doc__); sys.exit(2) | |
| incident = sys.argv[1] | |
| if len(sys.argv) >= 3: | |
| answer = Path(sys.argv[2]).read_text() | |
| else: | |
| answer = sys.stdin.read() | |
| task_dir = REPO / "tasks" / incident | |
| if not (task_dir / "template.sol").exists(): | |
| print(json.dumps({"incident": incident, "score": 0, | |
| "reason": f"unknown task (no tasks/{incident}/template.sol)"})) | |
| sys.exit(2) | |
| # tier only changes template selection for the "scratch" variant (not shipped | |
| # in v0.1); every v0.1 task uses template.sol, so tier is informational. | |
| ref = json.dumps({"incident": incident, "registry": f"tasks/{incident}", | |
| "tier": "scaffolded"}) | |
| if "```" not in answer: | |
| answer = "```solidity\n" + answer + "\n```" | |
| score, info = fv.evaluate_foundry(answer, ref) | |
| print(json.dumps({"incident": incident, "score": score, | |
| "reason": info["reason"][:300]})) | |
| if __name__ == "__main__": | |
| main() | |