Ameame1002/ToolsBench

TEB — Tool-use Efficiency Benchmark, frozen release best_20261111 v2

v2 (2026-11-13): same 185 tasks, answers and scoring; web_search and image_search are no longer offered to models (run protocol v3), exactly as in the maintainers' runs.

**中文使用说明：README_zh.md**。简单说就是：下载整个目录，把 API key 写进 keys.env，用 root 在目录里启动 Codex， 对它说"按 AGENTS.md 运行，要测的模型是 …"。它会依次完成检查、安装、验收（不调模型）、预检、正式运行和打包结果。

185 tasks (162 long-horizon, 23 short) in which a model works with real files and 31 tools, scored on the correctness of its answer and on how many of its tool calls made progress recognised by each task's evidence graph. Task files, scoring and labelling are frozen (hashes in docs/FREEZE.json inside the code archive).

Files

file	content
README_zh.md	中文使用说明（操作步骤、时间、要发回的内容）
AGENTS.md	the procedure for a coding agent (or a person): every step, its success criterion and failure handling
teb.sh	one command per step: check, install, nettest, accept, provider, preflight, run, status, stop, package
install.sh	verifies the archives, unpacks the code into ./teb and runs its setup
teb_release_best_20261111.tar.zst	code + 185 tasks (runner, tool server sources, launcher, docs)
teb_images_best_20261111.tar.zst	every Docker image (one docker load; exact image IDs of the release)
teb_tool_weights_best_20261111.tar.zst	weights of the tool server's local models (Whisper, bge-m3, DePlot, Grounding DINO, MacBERT, beat_this); optional
SHA256SUMS	checksums

Using a coding agent (Codex, Claude Code) — recommended

Requirements: Linux x86_64, root, Docker Engine + Compose v2 with the classic image store (check explains how to switch if Docker uses the containerd store), network access, 40 GB free on Docker's data root and 25 GB in the download folder. No GPU.

1. Download the whole dataset folder:

   hf auth login
   hf download Ameame1002/ToolsBench --repo-type dataset --local-dir teb_download
   

2. Put the API keys of the models to evaluate into teb_download/keys.env (never paste keys into the chat):

   cat > teb_download/keys.env <<'EOF'
   OPENAI_API_KEY=...
   ANTHROPIC_API_KEY=...
   GOOGLE_GENERATIVE_AI_API_KEY=...
   EOF
   chmod 600 teb_download/keys.env
   

3. As root, start the agent in that folder with permission to use Docker and the network, e.g. Codex:

   cd teb_download && codex --sandbox danger-full-access
   

4. Tell it, for example:

   Follow AGENTS.md. Evaluate: OpenAI <model-id>, Anthropic <model-id>, Google <model-id>. Docker on this machine may use CPUs. Report after the acceptance test and the preflights, then run the full benchmark and package the results.

AGENTS.md is the complete procedure (check → install → network check of web_fetch / poi_search → acceptance test without a model → configure models → 2-task preflight per model → full run in the background with status checks and retries → package). teb.sh implements each step as one command; long steps run in the background (bg / poll), so an agent session can end and a new one continues.

Install by hand

bash teb.sh check                          # as root
TEB_CPUS=<cpus> bash teb.sh install        # verifies, unpacks ./teb, loads images, starts the tool server
bash teb.sh nettest                        # live web_fetch / poi_search check through the tool server: must PASS
                                           # (FAIL = containers have no internet: TEB_UPSTREAM_PROXY=<url> bash teb.sh nettest)
bash teb.sh accept --jobs 8                # acceptance test, no model, no API cost: must PASS
bash teb.sh provider openai <model-id>     # or anthropic / google / compat (keys from env variables)
bash teb.sh preflight openai/<model-id>    # 2 short tasks: must PASS
bash teb.sh run r1 openai/<model-id>=4 --jobs 8
bash teb.sh status r1                      # until "VERDICT: COMPLETE" (INCOMPLETE -> fix the cause, bash teb.sh retry r1)
bash teb.sh package r1                     # r1.tar.gz to send back

Details: teb/README.md; 中文说明：teb/docs/QUICKSTART_zh.md。