ClawBench (nearai-bench packaging)

A flat, self-contained repackaging of ClawBench — 319 agent tasks across 35 domains, difficulty levels L1–L4. Task content, environments and verifiers are unmodified, so scores stay comparable to the upstream ClawBench leaderboard.

Why this exists

Upstream ships a git repo of nested task directories (tasks/<domain>/<task>/{task.toml,instruction.md,environment/,verifier/,solution/}). Cloning that per worker is wasteful for an eval/RL harness. Here each task is one row, with its three directory payloads as deterministic base64 tar.gz blobs.

from datasets import load_dataset
ds = load_dataset("NEAR-AI/clawbench", split="train")

Columns

Column	Type	Notes
task_id	string	Stable task id = the upstream task directory name (e.g. acct-001-journal-entries)
upstream_id	string	The id field inside task.toml. Often a short form (sec-001) that is not unique across domains — prefer task_id
task_path	string	<domain>/<task_id>, the task's path under upstream tasks/
title	string	Human-readable title
domain	string	One of 34 domains (email, security, multi-agent, …)
level	string	L1–L4 difficulty
track	string	foundation | subject-matter; empty when upstream omits it
description	string	One-line task description, when upstream supplies one
timeout	int64	Upstream per-task budget (seconds)
skills_allowed	bool	Whether the task permits skill creation/reuse
tags	string (JSON)	Upstream tag list
capabilities	string (JSON)	e.g. ["tool-use"] or ["file-read","file-write"] — see note below
capability_types	string (JSON)	e.g. ["reasoning","tool-use"]
required_actions	string (JSON)	e.g. ["file-read","data-processing","file-write"]
instruction	string	Verbatim instruction.md — the agent-facing prompt
task_toml	string	Verbatim task.toml
environment_tar	string	base64(tar.gz) of environment/ — setup.sh plus any data/ seed files
verifier_tar	string	base64(tar.gz) of verifier/ (pytest test_output.py) plus a bundled conftest.py
solution_tar	string	base64(tar.gz) of solution/ — the reference solve.sh

Note on the two upstream task.toml shapes

Upstream is not uniform: 65 tasks nest their metadata under a [task] table, the other 254 put the same keys at the top level — and the two shapes carve up capabilities differently (capabilities + required_actions vs. capabilities + capability_types). The flattened columns above normalize both, and task_toml always holds the verbatim original. If you parse task_toml yourself, handle both shapes or you will silently blank the metadata of 80% of the suite.

Running a task

import base64, io, subprocess, tarfile, tempfile, pathlib

def untar(b64, dest):
    if not b64: return
    dest.mkdir(parents=True, exist_ok=True)
    with tarfile.open(fileobj=io.BytesIO(base64.b64decode(b64)), mode="r:gz") as t:
        t.extractall(dest)

row = ds[0]
tmp = pathlib.Path(tempfile.mkdtemp())
untar(row["environment_tar"], tmp / "environment")
untar(row["verifier_tar"],   tmp / "verifier")
workspace = tmp / "workspace"; workspace.mkdir()

# 1. seed the workspace
subprocess.run(["bash", str(tmp / "environment/setup.sh"), str(workspace)], check=True)
# 2. give row["instruction"] to the agent, let it work in `workspace`
# 3. score with the upstream pytest verifier
subprocess.run(
    ["python", "-m", "pytest", "verifier/test_output.py", "--workspace", str(workspace), "-q"],
    cwd=tmp,
)

setup.sh takes the workspace directory as $1. Pass an absolute path — several scripts interpolate $1 into a heredoc that runs with a different cwd, so a relative path silently produces an empty workspace.

Scoring

The verifier is pytest. Each test carries an @pytest.mark.weight(n) marker (default 2.0); the task score is sum(weight of passing tests) / sum(all weights). conftest.py — bundled into every verifier_tar — provides both that marker and the --workspace option, so a verifier run needs nothing else from the upstream repo.

Two things worth knowing if you compare numbers:

• The verifier imports numpy/pandas for some domains. A verifier environment missing them yields false zeros rather than errors.
• The upstream leaderboard metric is a difficulty-weighted aggregate over a 5-dimension composite, not a flat mean of per-task scores.

A reference implementation lives in nearai/benchmarks at src/adapters/clawbench.rs.

⚠️ Contamination warning

solution_tar contains reference solutions. They are published upstream too, and they are needed for golden-validation (a correct harness must score 1.0 when the solution is applied and ~0.0 on an empty workspace) — but do not train on this column, and drop it before handing any of this to a model.

Provenance & license

• Upstream: https://github.com/claw-bench/claw-bench — Apache-2.0. Pinned commit 1fc25add8fe77aa498d58fb564ea91a87307da76.
• This repackaging: Apache-2.0, same terms. Task content unmodified; only the container format changed, plus a conftest.py copy bundled into each verifier_tar for self-containment.
• Packaged by: NEAR AI for nearai-bench.