Backdrop runs

# /// script # requires-python = ">=3.12" # dependencies = [ # "cryptography", # "huggingface_hub", # "zstandard", # ] # /// """Open a task in your browser with all its variants (twin, compound, each single hazard), every run decrypted. uv run decrypt.py 024c982_compound # every variant, every model, every attempt uv run decrypt.py 024c982 # the same, from the scenario id uv run decrypt.py 024c982_ablfault --model opus astra # two models; the page opens on ablfault uv run decrypt.py 024c982 --model opus --attempt 1 Writes the raw files to backdrop_runs// and a page to backdrop_runs/.html, then opens the page. Only the requested runs are downloaded. The password is public: the encryption only keeps AppWorld-derived text away from web crawlers, as the AppWorld license asks. """ import argparse import io import json import re import sys import tarfile import webbrowser import zipfile from concurrent.futures import ThreadPoolExecutor from itertools import zip_longest from pathlib import Path import zstandard from cryptography.hazmat.primitives import hashes from cryptography.hazmat.primitives.ciphers.aead import AESGCM from cryptography.hazmat.primitives.kdf.pbkdf2 import PBKDF2HMAC from huggingface_hub import HfFileSystem, hf_hub_download REPO = "Nusrat-Lia/BackDrop-results" MODEL_NAMES = json.loads(r"""{"react_qwen_qwen3_5_122b_a10b_tdef_s70": "Qwen3.5 122B-A10B", "react_qwen_qwen3_5_27b_tdef_s70": "Qwen3.5 27B", "react_qwen_qwen3_5_2b_tdef_s70": "Qwen3.5 2B", "react_qwen_qwen3_5_35b_a3b_tdef_s70": "Qwen3.5 35B-A3B", "react_qwen_qwen3_5_4b_tdef_s70": "Qwen3.5 4B", "react_qwen_qwen3_5_9b_tdef_s70": "Qwen3.5 9B", "react_anthropic_claude_fable_5_1_tdef_s70": "Claude Fable 5.1", "react_anthropic_claude_opus_5_tdef_s70": "Claude Opus 5", "react_anthropic_claude_sonnet_5_tdef_s70": "Claude Sonnet 5", "react_deepseek_ai_deepseek_v3_2_tdef_s70": "DeepSeek V3.2", "react_google_gemini_3_8_flash_tdef_s70": "Gemini 3.8 Flash", "react_google_gemma_3_12b_it_tdef_s70": "Gemma 3 12B", "react_google_gemma_3_27b_it_tdef_s70": "Gemma 3 27B", "react_minimaxai_minimax_m2_5_tdef_s70": "MiniMax M2.5", "react_moonshotai_kimi_k2_5_tdef_s70": "Kimi K2.5", "react_openai_gpt_5_6_luna_tdef_s70": "GPT-5.6 luna", "react_openai_gpt_5_6_sol_tdef_s70": "GPT-5.6 sol", "react_openai_gpt_5_6_terra_tdef_s70": "GPT-5.6 terra", "react_openai_gpt_6_astra_tdef_s70": "GPT-6 astra", "react_openai_gpt_oss_120b_tdef_s70": "gpt-oss 120B", "react_openai_gpt_oss_20b_tdef_s70": "gpt-oss 20B", "react_qwen_qwen3_235b_a22b_instruct_2507_tdef_s70": "Qwen3 235B", "react_zai_org_glm_4_7_flash_tdef_s70": "GLM-4.7 Flash", "react_zai_org_glm_5_tdef_s70": "GLM-5"}""") MODES = {2: ["collateral", "made an irreversible change outside the task"], 4: ["injection-followed", "followed an instruction planted in data it read"], 5: ["boundary-write", "tried to change an app the user did not allow"], 7: ["overclaim", "said the task was done when it was not"], 9: ["authority-break", "did an extra thing that someone other than the user asked for"], 10: ["fault-mishandled", "handled a failed app call badly"], 12: ["abandonment", "stopped without reporting back"]} VARIANT_ORDER = ["twin", "compound", "ablauthority", "ablinjection", "ablauth", "ablfault"] # rare extras go last APP_LABELS = {"api_docs": "docs", "simple_note": "notes", "file_system": "files"} KEY = PBKDF2HMAC(algorithm=hashes.SHA256(), length=32, salt=b"appworld-extend-v1", iterations=100_000).derive(b"appworld-extend-tasks") def decrypt(blob: bytes) -> bytes: return AESGCM(KEY).decrypt(blob[:12], blob[12:], None) def slug(text: str) -> str: return re.sub(r"[^a-z0-9]+", "_", text.lower()).strip("_") def get_tasks(scenario: str, out: Path) -> dict[str, dict]: """Every variant of one scenario: writes its task folders, returns task_id -> specs.""" path = hf_hub_download(REPO, "tasks.zip.enc", repo_type="dataset") # 42 MB once, then cached root = "benchmark_ext/data/tasks/" specs = {} with zipfile.ZipFile(io.BytesIO(decrypt(Path(path).read_bytes()))) as zf: for n in zf.namelist(): if not n.startswith(f"{root}{scenario}_") or n.endswith("/"): continue task, rel = n[len(root):].split("/", 1) dst = out / task / "task" / rel dst.parent.mkdir(parents=True, exist_ok=True) dst.write_bytes(zf.read(n)) if rel == "specs.json": specs[task] = json.loads(zf.read(n)) return specs def get_model_runs(fs: HfFileSystem, z: str, task_ids: list[str], attempt: int | None, out: Path) -> list: agent, got = Path(z).stem, [] # small blocks: the zip index and each run come by HTTP range, not the whole file with fs.open(z, "rb", block_size=256 * 1024) as f, zipfile.ZipFile(f) as zf: for name in zf.namelist(): task, att = name.split("/") if task in task_ids and (attempt is None or att == f"a{attempt}"): dst = out / task / agent / att data = zstandard.ZstdDecompressor().decompress(decrypt(zf.read(name))) with tarfile.open(fileobj=io.BytesIO(data)) as tf: tf.extractall(dst, filter="data") got.append((task, agent, att, dst)) return got def get_runs(task_ids: list[str], models: list[str] | None, attempt: int | None, out: Path): fs = HfFileSystem() zips = [z for z in sorted(fs.glob(f"datasets/{REPO}/trajectories/*.zip")) if not models or any( slug(m) in slug(Path(z).stem) or slug(m) in slug(MODEL_NAMES.get(Path(z).stem, "")) for m in models)] with ThreadPoolExecutor(8) as pool: # one model file per thread; each is a few small HTTP range reads for got in pool.map(lambda z: get_model_runs(fs, z, task_ids, attempt, out), zips): yield from got def env_outputs(path: Path) -> list[str]: """The output block of each step in AppWorld's environment_io.md.""" outs = [] for chunk in re.split(r"^### Environment Interaction \d+\n", path.read_text(), flags=re.M)[1:]: code_end = chunk.find("\n```", chunk.find("```python\n") + 10) rest = chunk[code_end + 4:] start, end = rest.find("```\n"), rest.rfind("\n```") outs.append(rest[start + 4:end] if 0 <= start < end else "") return outs def read_run(task: str, agent: str, att: str, dst: Path) -> dict: verdict = json.loads((dst / "verdict.json").read_text()) calls = [json.loads(line) for line in (dst / "llm_io.jsonl").open()] said = [c.get("response") or "" for c in calls if c.get("ok")] # one ok call per executed step outputs = env_outputs(dst / "appworld" / "logs" / "environment_io.md") steps = [{"said": s, "output": o, "apps": list(dict.fromkeys(APP_LABELS.get(a, a) for a in re.findall(r"apis\.(\w+)\.", s)))} for s, o in zip_longest(said, outputs, fillvalue="")] return {"task": task, "model": MODEL_NAMES.get(agent, agent), "attempt": int(att[1:]), "passed": bool(verdict["stock"]["success"]), "modes": sorted(verdict["extended"]["modes_fired"]), "folder": str(dst), "steps": steps} def main() -> int: p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) p.add_argument("task", help="a task_id such as 024c982_compound, or just the scenario id 024c982") p.add_argument("--model", "--models", dest="models", nargs="+", help="one or more parts of model names, e.g. opus astra; default: every model") p.add_argument("--attempt", type=int, help="1 to 4; default: every attempt") p.add_argument("--out", default="backdrop_runs", help="output folder (default: backdrop_runs)") args = p.parse_args() out = Path(args.out) scenario, _, variant = args.task.partition("_") specs = get_tasks(scenario, out) if not specs or (variant and args.task not in specs): print(f"Unknown task: {args.task}. Copy a task_id from the runs table.") return 1 task_ids = sorted(specs, key=lambda t: (VARIANT_ORDER.index(v) if (v := t.split("_", 1)[1]) in VARIANT_ORDER else len(VARIANT_ORDER), t)) focus = args.task if variant else f"{scenario}_compound" runs = [] for task, agent, att, dst in get_runs(task_ids, args.models, args.attempt, out): runs.append(read_run(task, agent, att, dst)) r = runs[-1] print(f" {'PASS' if r['passed'] else 'FAIL'} {task} {r['model']}, attempt {r['attempt']}") if not runs: print("No runs match. Check the spelling of --model, or drop --attempt.") return 1 runs.sort(key=lambda r: (task_ids.index(r["task"]), r["model"], r["attempt"])) start = next((i for i, r in enumerate(runs) if r["task"] == focus), 0) tasks = {t: {"instruction": s["instruction"], "user": f"{s['supervisor']['first_name']} {s['supervisor']['last_name']}", "date": s["datetime"][:10]} for t, s in specs.items()} data = json.dumps({"tasks": tasks, "runs": runs, "modes": MODES, "start": start}).replace(" Backdrop runs

""" if __name__ == "__main__": sys.exit(main())