From eb850a2887aca7c05aef0a4b0aaab26662e229f6 Mon Sep 17 00:00:00 2001
From: Ylr9933 <2835703479@qq.com>
Date: Wed, 23 Sep 2026 01:46:30 +0800
Subject: [PATCH] Add Terminal-Bench adapter (terminal-bench-core==0.1.1)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Terminal-Bench (laude-institute/terminal-bench) is the docker-per-task
resolution benchmark. Adds it as a Benchmark adapter in
experiments/benchmark/terminal_bench/, following alfworld's shape.

Version lock — terminal-bench-core==0.1.1 (registry.json published entry):
  commit 91e10457b5410f16c44364da1a34cb6de8c488a5
  branch dataset/terminal-bench-core/v0.1.x
  task_id_subset 80 ids (.base/.easy/.hard variants) -> 70 unique task dirs
(repo has no git tags; PyPI terminal-bench-core 404; 0.1.1 is the latest
registerable release. main HEAD drifts and is not reproducible, so not used.)

Adapter surface:
- tasks(): 70 curated dirs; SUBSET_BASES embedded as a module tuple
  (clone-only source, no dataset-side file). Sampling rewoo_draw(seed).
- on_task(): builds + runs the per-task container (base image
  ghcr.io/laude-institute/t-bench/python-3-13, multi-arch); mounts
  run-tests.sh + tests/ into /app:ro; the lone `exec` tool docker-exec's in.
- score(): runs the official run-tests.sh (which runs `uv run pytest tests/`)
  in the agent-modified container end-state; correctness = pytest pass
  (the dataset's own judge, not a reimplementation).
- downloads(): pins the codeload tarball @ 91e10457 with the registry note.
- Registered in run.py BENCHMARK_FACTORIES; DOWNLOADS.md regenerated to cover it.
- .gitignore: ignore .venv/.

Gates:
- `python -m experiments.scripts.datasets --write` ok
- `python -m experiments.scripts.check`: 16 benches · 10 arms · 0 runs, all pass
- docker smoke on this arm64 mac: base image builds natively in 27s (no
  emulation); exec + mounted run-tests path works; end-to-end
  react × DeepSeek 1 task finishes through official pytest judging.

Known limits:
- ~10 heavy tasks (qemu kernel / hf-model / pytorch-model-cli ...) may
  not build on arm64 mac; recorded as env_unavailable. ~60/70 run here;
  all 70 should run on x86 linux.
- per-task ~4-7 min (apt/uv install pytest inside score can't be cached);
  70 tasks ≈ 3-6 h.
- .easy/.hard variants not split; each unique dir runs once.
---
 .gitignore                                    |   3 +
 .../benchmark/terminal_bench/__init__.py      |   7 +
 .../terminal_bench/terminal_bench.py          | 338 ++++++++++++++++++
 experiments/dataset/DOWNLOADS.md              |   1 +
 experiments/scripts/run.py                    |   3 +
 5 files changed, 352 insertions(+)
 create mode 100644 experiments/benchmark/terminal_bench/__init__.py
 create mode 100644 experiments/benchmark/terminal_bench/terminal_bench.py

diff --git a/.gitignore b/.gitignore
index ef477b7..c90fd7a 100644
--- a/.gitignore
+++ b/.gitignore
@@ -39,6 +39,9 @@ __pycache__/
 .pytest_cache/
 .ruff_cache/
 
+# 虚拟环境（每人本地建、不进仓库）
+.venv/
+
 # ★ 论文插图：**版权不属于我们，不进公开仓库。**
 #   arXiv 的授权因文而异（有的 CC-BY 可以带署名转载，有的不是），
 #   而这些图是**我们的读书材料**，不是代码的一部分。
diff --git a/experiments/benchmark/terminal_bench/__init__.py b/experiments/benchmark/terminal_bench/__init__.py
new file mode 100644
index 0000000..7ff532a
--- /dev/null
+++ b/experiments/benchmark/terminal_bench/__init__.py
@@ -0,0 +1,7 @@
+from experiments.benchmark.terminal_bench.terminal_bench import (  # noqa: F401
+    NeedsDocker,
+    TB_COMMIT,
+    TerminalBench,
+)
+
+__all__ = ["TerminalBench", "NeedsDocker", "TB_COMMIT"]
diff --git a/experiments/benchmark/terminal_bench/terminal_bench.py b/experiments/benchmark/terminal_bench/terminal_bench.py
new file mode 100644
index 0000000..8802923
--- /dev/null
+++ b/experiments/benchmark/terminal_bench/terminal_bench.py
@@ -0,0 +1,338 @@
+"""Terminal-Bench —— **每题一个 Docker 容器环境，官方 pytest 判分。**
+
+## 出处
+
+| | |
+|---|---|
+| **论文** | *Terminal-Bench: A Benchmark for End-to-End Tasks in Real Terminal Environments*. arXiv:2601.11868 |
+| **仓库** | `github.com/laude-institute/terminal-bench`（HEAD 没固定;改用 registry 给的 lock 版本）|
+| **dataset** | **`terminal-bench-core==0.1.1`**（registry.json publish 的 latest release entry）<br>commit `91e10457b5410f16c44364da1a34cb6de8c488a5`<br>branch `dataset/terminal-bench-core/v0.1.x`<br>`task_id_subset` 80 个 id(含 `.easy`/`.hard` 变体) → 去点 **70 个 unique task dir**<br>每 task: `task.yaml` + `Dockerfile` + `docker-compose.yaml` + `run-tests.sh` + `tests/` |
+
+## ★ 这是个**交互式容器环境**，形状和 ALFWorld 那类一样
+
+每题是一个 docker 容器（`command: sleep infinity`），agent 通过 `docker exec` 在容器里跑命令来「做题」，
+判分是**在 agent 改过的容器终态上跑官方 `run-tests.sh`**（它跑 `uv run pytest tests/`）。
+「对不对」是 **pytest 的通过/失败** —— 这是官方 judge，不自己写判分器。
+
+## ★ 判分口径（照学长群里第 ② 条：调官方别自己写）
+
+`parser_name: pytest` 的 task 判据 = **`pytest` 的 pass/fail**。我们不重写 parser：
+跑 `docker exec <container> sh -c 'bash ./run-tests.sh'`，读 pytest 的退出码与 LS 输出里的
+`failed` 计数。correctness = `exit==0 且无 failed`。该任务的官方判据就是 pytest，所以这一步
+不是「自己抄 parser」，是「跑官方给的 pytest 并读它的结论」。
+
+## ⚠️ 必须记住的坑（读原仓库 + CLAUDE.md 读出来）
+
+- **镜像每题现场 build**（`Dockerfile` 在 task 目录里）。100 题里大部分 task 各自一个镜像，
+  这是跑这一项的主要时间成本（不是 deepseek 的钱，是 docker build）。
+- **arm64 mac 上部分 task 跑不了**（如 pin 了 win32/x86 的 `3d-model-format-legacy`）。
+  `docker build` 失败的那题，判分里如实记 `env_unavailable`，不静默当对。
+- **不依赖 `terminal-bench` pip 包**（它要 Python 3.13 + postgres/supabase，太重，且它带自己的
+  agent，我们要用自己的 `react` 臂）。这里直接用 `docker` CLI 管 build/exec/run-tests。
+
+## ★ 抽样用 `rewoo_draw`（群里第 ③ 条点名要的）
+
+  --seed 决定我们从 79 题里抽哪一批，全项目同一种做法。
+"""
+
+from __future__ import annotations
+
+import re
+import shutil
+import subprocess
+import uuid
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Iterator, Sequence
+
+import yaml
+
+from experiments.core.download import DATASET_DIR, DownloadSpec
+from experiments.core.types import Judgment, Task, Tool, Trajectory
+
+#: Terminal-Bench **published dataset** `terminal-bench-core==0.1.1`
+#:   —— 不跟 main（main 一夜会换 task 集，不可复现）。registry.json 里登记的
+#:   最 release 的 lock 版本: commit + branch + task_id_subset 三件全。
+TB_COMMIT = "91e10457b5410f16c44364da1a34cb6de8c488a5"
+TB_VERSION = "0.1.1"
+TB_BRANCH = "dataset/terminal-bench-core/v0.1.x"
+
+#: `v0.1.1` 分支下 task 定义在 `tasks/<task>/`（main 后来才改名 `original-tasks/`）。
+TB_ROOT = DATASET_DIR / "terminal_bench" / "v0.1.1"
+TASKS_DIR = TB_ROOT / "tasks"
+#: v0.1.1 里 `datasets/terminal-bench-core-v0.yaml` 的 `task_ids`（80 个，含
+#: `.base`/`.easy`/`.hard` 变体）**去后缀**的 70 个 unique task dir ——
+#: 圈定的 published subset。不跑分支全集 86 个，只跑这 70。
+#: （变体 .easy/.hard 是同一 task dir 不同难度，paper 一同报；我们这版每道跑一次，
+#:  70 是 published subset 的 unique base，可复现。）
+SUBSET_BASES: tuple[str, ...] = (
+    'blind-maze-explorer-5x5', 'blind-maze-explorer-algorithm', 'build-initramfs-qemu', 'build-linux-kernel-qemu',
+    'build-tcc-qemu', 'cartpole-rl-training', 'chess-best-move', 'conda-env-conflict-resolution',
+    'configure-git-webserver', 'count-dataset-tokens', 'crack-7z-hash', 'create-bucket',
+    'cron-broken-network', 'csv-to-parquet', 'decommissioning-service-with-sensitive-data', 'download-youtube',
+    'eval-mteb', 'extract-moves-from-video', 'extract-safely', 'fibonacci-server',
+    'fix-git', 'fix-pandas-version', 'fix-permissions', 'get-bitcoin-nodes',
+    'git-multibranch', 'git-workflow-hack', 'gpt2-codegolf', 'grid-pattern-transform',
+    'hello-world', 'heterogeneous-dates', 'hf-model-inference', 'incompatible-python-fasttext',
+    'intrusion-detection', 'jupyter-notebook-server', 'modernize-fortran-build', 'new-encrypt-command',
+    'nginx-request-logging', 'oom', 'openssl-selfsigned-cert', 'organization-json-generator',
+    'password-recovery', 'path-tracing', 'path-tracing-reverse', 'play-zork',
+    'polyglot-c-py', 'polyglot-rust-c', 'processing-pipeline', 'prove-plus-comm',
+    'pytorch-model-cli', 'qemu-alpine-ssh', 'qemu-startup', 'raman-fitting',
+    'reshard-c4-data', 'run-pdp11-code', 'sanitize-git-repo', 'security-vulhub-minio',
+    'simple-sheets-put', 'simple-web-scraper', 'solana-data', 'sqlite-db-truncate',
+    'sqlite-with-gcov', 'super-benchmark-upet', 'swe-bench-astropy-1', 'swe-bench-astropy-2',
+    'swe-bench-fsspec', 'swe-bench-langcodes', 'tmux-advanced-workflow', 'train-fasttext',
+    'vim-terminal-task', 'write-compressor',
+)
+IMAGE_PREFIX = "jevloop-tb"
+
+EXEC_TOOL = Tool(
+    name="exec",
+    description=("Run a shell command in the task's container. The command runs in the task's "
+                 "working directory; you see its stdout/stderr. Use this to inspect and modify "
+                 "the environment to complete the task."),
+    parameters={"type": "object",
+                "properties": {"command": {"type": "string"}},
+                "required": ["command"]},
+)
+
+
+class NeedsDocker(RuntimeError):
+    """要用容器（exec/score）但 docker 不可用，或这一题的镜像在本机 build 不起来。"""
+
+
+def _run(cmd: list[str], *, timeout: int = 1800, **kw) -> subprocess.CompletedProcess:
+    """跑一条命令，超时和失败都抛清楚（不把 stderr 当 stdout 吞）。"""
+    return subprocess.run(cmd, capture_output=True, text=True, timeout=timeout, **kw)
+
+
+@dataclass
+class TerminalBench:
+    """Terminal-Bench v0.1.1 published dataset（70 unique task dirs）。`split` 只有 `test`。"""
+
+    name: str = "terminal-bench"
+    dataset_version: str = f"terminal-bench-core=={TB_VERSION}"
+    data_dir: Path = field(default_factory=lambda: DATASET_DIR / "terminal_bench")
+    split: str = "test"
+    limit: int | None = None
+    seed: int = 0
+    #: agent 单题最多步数。TB task 有的要很久，runner 的 `--max-steps` 是默认 20，可经命令行覆盖。
+    max_steps: int = 40
+    #: docker build 单题超时（秒）。legacy build 慢，给宽一点。
+    build_timeout_s: int = 1800
+    #: run-tests.sh 超时（秒）。
+    test_timeout_s: int = 600
+    exec_timeout_s: int = 120
+    #: 当前题。`score()` 在 agent 改过的**同一容器**终态上跑官方 `run-tests.sh`——
+    #   不同于 ALFWorld 的「回放」，TB 的判据是容器状态被 agent 改成什么样了。
+    _current: Task | None = field(default=None, init=False, repr=False)
+    #: 这一题起出来的容器名。每题一个独立容器。
+    _container: str | None = field(default=None, init=False, repr=False)
+
+    # ── 数据 ────────────────────────────────────────────────
+
+    def _task_dirs(self) -> list[Path]:
+        if not TASKS_DIR.is_dir():
+            raise FileNotFoundError(
+                f"{self.name}: 缺 {TASKS_DIR}\n"
+                f"  跑 `python3 -m experiments.scripts.datasets --fetch terminal-bench`"
+            )
+        all_dirs = sorted(d for d in TASKS_DIR.iterdir()
+                          if d.is_dir() and (d / "task.yaml").exists())
+        want = set(SUBSET_BASES)
+        kept = [d for d in all_dirs if d.name in want]
+        if len(kept) != len(SUBSET_BASES):
+            have = {d.name for d in kept}
+            missing = sorted(want - have)[:8]
+            raise ValueError(
+                f"{self.name}: subset 里有 {missing} 等 task 不在 {TASKS_DIR}。"
+                f"  → 数据版本和 subset 不一致？"
+            )
+        return kept
+
+    def tasks(self, *, split: str, limit: int | None, seed: int) -> Iterator[Task]:
+        """出题。只取 0.1.1 `task_id_subset` 的 70 个 unique base dir;抽样 `rewoo_draw(seed)`。"""
+        from experiments.benchmark.rewoo_port import rewoo_draw
+
+        _ = split or self.split  # 官方只有 `test`；不分 train/test
+        dirs = self._task_dirs()
+        for i in rewoo_draw(len(dirs), limit if limit is not None else self.limit, seed):
+            d = dirs[i]
+            meta = yaml.safe_load((d / "task.yaml").read_text(encoding="utf-8")) or {}
+            instruction = str(meta.get("instruction") or "").strip()
+            if not instruction:
+                continue
+            yield Task(
+                task_id=f"terminal-bench/{d.name}",
+                prompt=(instruction
+                        + "\n\nYou have a shell tool (`exec`). Inspect and modify the environment "
+                        "to complete the task; stop calling tools when finished."),
+                gold=None,  # TB 没有字符串金标——判据是 pytest 在容器终态的结果
+                oracle_context=None,
+                meta={
+                    "task_name": d.name,
+                    "difficulty": meta.get("difficulty"),
+                    "category": meta.get("category"),
+                    "parser": meta.get("parser_name"),
+                    "estimated_duration_sec": meta.get("estimated_duration_sec"),
+                    "task_dir": str(d),
+                    "answer_kind": "environment",
+                },
+            )
+
+    # ── 工具与环境 ────────────────────────────────────────────
+
+    def on_task(self, task: Task) -> None:
+        """runner 每题调一次。停下上一题的容器（每题独立容器），记这一题。"""
+        self._stop()
+        self._current = task
+        self._container = None
+
+    def _has_image(self, task_dir: Path) -> bool:
+        r = _run(["docker", "image", "inspect", f"{IMAGE_PREFIX}:{task_dir.name}"], timeout=10)
+        return r.returncode == 0
+
+    def _ensure_container(self) -> str:
+        """懒建这一题的镜像 + 容器。第一次 `tool_impls()` / `score()` 调时触发。"""
+        if self._container is not None:
+            return self._container
+        if self._current is None:
+            raise NeedsDocker(f"{self.name}: 没绑任务，先调 on_task(task)")
+        if not shutil.which("docker"):
+            raise NeedsDocker(f"{self.name}: 需要 `docker`（build/exec/run-tests）——没装")
+
+        task_dir = Path(self._current.meta["task_dir"])
+        image = f"{IMAGE_PREFIX}:{task_dir.name}"
+
+        if not self._has_image(task_dir):
+            br = _run(["docker", "build", "-t", image, "-f", "Dockerfile", "."],
+                      cwd=str(task_dir), timeout=self.build_timeout_s)
+            if br.returncode != 0:
+                # ★ 多半是这题在本机 build 不动（arm64 起不来 x86/win32 镜像）。
+                #   不是「task 不可解」，是「环境跑不了这题」——分开记。
+                tail = br.stderr[-800:] or br.stdout[-800:]
+                raise NeedsDocker(
+                    f"{self.name}: docker build 失败 ({task_dir.name})\n{tail}"
+                )
+        cont = f"jevloop-tb-{uuid.uuid4().hex[:10]}"
+        # 把 task dir 的 `run-tests.sh` + `tests/` 挂进容器 `/app`（base image 默认 WORKDIR）。
+        # ★ 大部分 Dockerfile **不 COPY 这两样**（broken-python 等只 `FROM base + RUN ...`），
+        #   Terminal-Bench 官方 harness 是靠绑定挂载把它们弄进容器的——我们手动 run 时
+        #   不挂就缺，判分找不到 `run-tests.sh`（实测）。`:ro` 让判分脚本和测试不被 agent 改动，
+        #   但镜像 /app 里的业务文件 agent 仍可写。
+        td = str(task_dir.resolve())
+        rr = _run(["docker", "run", "-d", "--name", cont, "--workdir", "/app",
+                   "-v", f"{td}/run-tests.sh:/app/run-tests.sh:ro",
+                   "-v", f"{td}/tests:/app/tests:ro",
+                   "-e", "TEST_DIR=./tests", image, "sleep", "infinity"],
+                  timeout=60)
+        if rr.returncode != 0:
+            raise NeedsDocker(f"{self.name}: docker run 失败\n{rr.stderr[-800:]}")
+        self._container = cont
+        return cont
+
+    def tools(self) -> Sequence[Tool]:
+        self._ensure_container()
+        return (EXEC_TOOL,)
+
+    def tool_impls(self) -> dict[str, Any]:
+        self._ensure_container()
+
+        def exec_(command: str = "", **kwargs: Any) -> str:
+            cmd = str(command or kwargs.get("arg") or "").strip()
+            if not cmd:
+                return "Error: empty command."
+            r = _run(["docker", "exec", self._container, "sh", "-c", cmd],
+                     timeout=self.exec_timeout_s)
+            out = r.stdout
+            if r.stderr:
+                out += (("\n[stderr]\n" + r.stderr) if r.stderr.strip() else "")
+            if r.returncode != 0:
+                out += f"\n[exit={r.returncode}]"
+            # 截一下，免得一条 ls -R 把 context 撑爆
+            return out[-8000:] if len(out) > 8000 else out
+
+        return {"exec": exec_}
+
+    # ── 判分 ────────────────────────────────────────────────
+
+    def score(self, task: Task, trajectory: Trajectory) -> Judgment:
+        """在 agent 改过的同一容器里跑官方 `run-tests.sh`，读 pytest 的结果。
+
+        correctness = run-tests.sh 退出 0 **且** 输出里没有 `failed` 计数 > 0。
+        pytest 自身就是这一题的官方判据，所以这里**不重写 parser**——
+        跑它给的脚本、读它给的结论。
+        """
+        # 即使 agent 一条 exec 没调也得能判分（必然失败，但要走完记账）
+        try:
+            self._ensure_container()
+        except NeedsDocker as exc:
+            return Judgment(correct=False, score=0.0, detail=str(exc),
+                             failure_class="env_unavailable")
+
+        task_dir = Path(task.meta["task_dir"])
+        r = _run(["docker", "exec", self._container, "sh", "-c", "bash ./run-tests.sh"],
+                 timeout=self.test_timeout_s)
+        out = (r.stdout or "") + (("\n[stderr]\n" + r.stderr) if (r.stderr or "").strip() else "")
+        failed = _count_failed(out)
+        passed = (r.returncode == 0) and failed == 0
+
+        detail = f"run-tests.sh exit={r.returncode}; pytest failed={failed}"
+        if not r.stdout and r.stderr:
+            detail = f"只输出到 stderr（exit={r.returncode}）：{r.stderr[-300:]}"
+        return Judgment(
+            correct=passed, score=1.0 if passed else 0.0,
+            detail=detail,
+            failure_class=None if passed else "task_not_completed",
+        )
+
+    def check(self, task: Task, answer: str) -> bool:
+        """Reflexion 的 Evaluator 要的成败信号 —— TB 的成败在容器状态，不在答案字符串。"""
+        raise NeedsDocker(
+            f"{self.name}: 成败由容器里的官方 pytest 判定，不是对答案字符串打分。"
+            f"用 score()（在容器跑 run-tests.sh），或让臂通过 session.check_answer 接上。"
+        )
+
+    # ── 下载 ────────────────────────────────────────────────
+
+    def downloads(self) -> Sequence[DownloadSpec]:
+        return [
+            DownloadSpec(
+                dataset=self.name,
+                kind="http",
+                # ★ registry.json 里登记的 v0.1.1: commit + branch 都写明
+                locator=(f"https://codeload.github.com/laude-institute/terminal-bench"
+                         f"/tar.gz/{TB_COMMIT}"),
+                files=("tasks/", "docker/", "registry.json"),
+                revision=TB_COMMIT,
+                size_hint="~14 MB（repo @ v0.1.1，不含 docker 镜像）",
+                note=(f"★ **terminal-bench-core=={TB_VERSION}** —— registry.json 里 publish "
+                      f"的 lock entry:commit {TB_COMMIT[:8]} / branch {TB_BRANCH}。"
+                      "解压到 `dataset/terminal_bench/v0.1.1/`（顶层名是 codeloud 的前缀）。"
+                      "docker 镜像**不在**这份里,每题 Dockerfile 现场 build。"
+                      "task_id_subset 80 个 id (含 .easy/.hard 变体) 的 unique base = 70 dir。"),
+            ),
+        ]
+
+    # ── 清理 ────────────────────────────────────────────────
+
+    def _stop(self) -> None:
+        if self._container:
+            _run(["docker", "stop", self._container], timeout=30)
+            _run(["docker", "rm", "-f", self._container], timeout=30)
+            self._container = None
+
+
+_FAILED_RE = re.compile(r"(\d+)\s+failed", re.IGNORECASE)
+
+
+def _count_failed(test_output: str) -> int:
+    """从 pytest 输出里抠 `failed` 计数；抠不到返回 -1（未知，但不当作 0 —— 见 gsm8k 的口径教训）。"""
+    m = _FAILED_RE.search(test_output)
+    if not m:
+        return 0 if test_output.strip() else -1
+    return int(m.group(1))
+
+
+__all__ = ["TerminalBench", "NeedsDocker", "TB_COMMIT", "TASKS_DIR", "EXEC_TOOL"]
diff --git a/experiments/dataset/DOWNLOADS.md b/experiments/dataset/DOWNLOADS.md
index f409801..45a4181 100644
--- a/experiments/dataset/DOWNLOADS.md
+++ b/experiments/dataset/DOWNLOADS.md
@@ -32,6 +32,7 @@
 | `tau2-bench` | http | `https://codeload.github.com/sierra-research/tau2-bench/tar.gz/refs/tags/v0.1.0` | `9` 个 | 37199f36924c | ~56 MB（整个仓库；本数据集约 25 MB） | ★ **必须钉 v0.1.0** —— 仓库 HEAD 已经是 τ³-bench v1.0.1,而它自己的 README 写着 `<1.0.1` 的结果不能和 `>=1.0.1` 比。
   ★ 而 v0.1.0 的 `tasks.json` **就是 base 划分**（278 条）;HEAD 上多了 `split_tasks.json`,base 要从里面挑 —— 换版本连哪 278 条都会变。
   ★ `test` 那一列在排行榜口径下要跑 **pass^k = 4 次/任务**（`comb(success,k)/comb(trials,k)`） |
+| `terminal-bench` | http | `https://codeload.github.com/laude-institute/terminal-bench/tar.gz/91e10457b5410f16c44364da1a34cb6de8c488a5` | `3` 个 | 91e10457b5410f16c44364da1a34cb6de8c488a5 | ~14 MB（repo @ v0.1.1，不含 docker 镜像） | ★ **terminal-bench-core==0.1.1** —— registry.json 里 publish 的 lock entry:commit 91e10457 / branch dataset/terminal-bench-core/v0.1.x。解压到 `dataset/terminal_bench/v0.1.1/`（顶层名是 codeloud 的前缀）。docker 镜像**不在**这份里,每题 Dockerfile 现场 build。task_id_subset 80 个 id (含 .easy/.hard 变体) 的 unique base = 70 dir。 |
 | `triviaqa` | hf-dataset | `trivia_qa:rc.nocontext` | — | ⚠️ **main**（未钉） | ~700 MB（138,384 + 17,944 + 17,210） | ★ config 必须是 `rc.nocontext` —— ReWOO 用的就是它。★ 答案带**别名集**,官方口径与 ReWOO 口径不同,见本文件头部。`revision` 未钉,已用 fingerprint 记进 dataset_version |
 
 ## 怎么用
diff --git a/experiments/scripts/run.py b/experiments/scripts/run.py
index b058d11..8e6eb1e 100644
--- a/experiments/scripts/run.py
+++ b/experiments/scripts/run.py
@@ -39,6 +39,7 @@
 # ★ 三个 BigBench 任务**共用一份 ReWOO tarball** —— 一次下载抽三个 CSV。
 from experiments.benchmark import bigbench as bigbench_bench  # noqa: F401
 from experiments.benchmark import alfworld as alfworld_bench  # noqa: F401
+from experiments.benchmark import terminal_bench as terminal_bench_bench  # noqa: F401
 from experiments.benchmark import fever as fever_bench  # noqa: F401
 from experiments.benchmark import gsm8k as gsm8k_bench  # noqa: F401
 from experiments.benchmark import hotpotqa as hotpotqa_bench  # noqa: F401
@@ -153,6 +154,8 @@ def build_model(args: argparse.Namespace) -> ModelClient:
     sotuqa_bench.SotuQa.name: sotuqa_bench.SotuQa,
     # ★ 交互式环境:需要 textworld,而且要用 `on_task` 绑定每题的环境
     alfworld_bench.AlfWorld.name: alfworld_bench.AlfWorld,
+    # ★ Terminal-Bench: 每题一个 docker 容器，判分跑官方 pytest（见 benchmark/terminal_bench/terminal_bench.py 头部）
+    terminal_bench_bench.TerminalBench.name: terminal_bench_bench.TerminalBench,
     # ★ 两方任务:还需要一个 LLM 用户模拟器（见它文件头的说明）
     tau2_bench.Tau2Bench.name: tau2_bench.Tau2Bench,
 }