task_id stringlengths 20 20 | name stringlengths 20 20 | domain stringclasses 9
values | skill_type stringclasses 29
values | task_complexity stringclasses 4
values | command_complexity stringclasses 3
values | scenario stringclasses 73
values | language stringclasses 8
values | anchor stringclasses 38
values | verifier_kind stringclasses 1
value | fixture_kind stringclasses 7
values | corpus_kind stringclasses 2
values | base_image stringclasses 1
value | ablation stringclasses 1
value | description stringlengths 967 7.14k | truth stringlengths 777 11.8k | primitive_skills listlengths 3 5 | metadata_json stringlengths 3.71k 17.7k | setup stringlengths 169 11.5k | container_def stringlengths 233 12.3k | test_initial_state stringlengths 271 6.12k | test_final_state stringlengths 705 15.6k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
task_000000_09a5f8a7 | task_000000_09a5f8a7 | software_engineering | Data Processing | intricate task (multi-stage workflow combining (a) environment setup or package configuration, (b) primary implementation across multiple files or languages, (c) iterative refinement against a quantitative or adversarial verifier, and (d) a final integration step. Expect 30-60 commands.) | bash and code (shell commands and writing/running scripts in Python, Perl, Ruby, etc.) | release manager preparing deployments | C | a Node.js project whose npm install fails due to conflicting peer deps | exact_text | multi_service_compose | rl_v2 | intricate | no_verifiers | You are the release manager for a high-throughput data processing pipeline. Our integration tests are failing after a recent protocol upgrade, and the upcoming deployment is blocked.
The system operates via a multi-service setup located in `/home/user/app/`:
1. **Telemetry Gateway (`/home/user/app/gateway`)**: A Go-b... | The setup requires creating the environment and buggy code files before the agent starts.
Run the following bash commands to prepare the `/home/user/app` directory.
```bash
mkdir -p /home/user/app/{c_src,gateway,tester,lib}
# 1. Create the initial old shared library (v1) to simulate the environment state
cat << 'EOF... | [
"Shared library and ABI management",
"WebSocket communication",
"Go concurrency patterns (goroutines, channels)",
"Character and data encoding"
] | {"ablation": "no_verifiers", "anchor": "a Node.js project whose npm install fails due to conflicting peer deps", "base_image": "intricate", "command_complexity": "bash and code (shell commands and writing/running scripts in Python, Perl, Ruby, etc.)", "corpus_kind": "rl_v2", "description": "You are the release manager ... | #!/bin/bash
set -euo pipefail
mkdir -p /home/user/.setup_tmp
export PYTHONPATH=/home/user/.local/lib/python3/dist-packages:${PYTHONPATH:-}
# Install Go 1.20
rm go1.20.14.linux-amd64.tar.gz
mkdir -p /home/user/app/c_src
mkdir -p /home/user/app/gateway
mkdir -p /home/user/app/tester
mkdir -p /home... | Bootstrap: docker
From: ubuntu:22.04
%environment
export PATH=/usr/local/go/bin:$PATH
%files
/gpfs/scrubbed/hamishiv/tmax-ablation-generation-20261003/runs/no_verifiers/chunks/rl_v2_0003/task_000000_09a5f8a7/fixtures/compose/nginx.conf /app/fixtures/compose/nginx.conf
/gpfs/scrubbed/hamishiv/tmax-ablation... | # test_initial_state.py
import os
import stat
import pytest
def test_directories_exist():
dirs = [
"/home/user/app",
"/home/user/app/c_src",
"/home/user/app/gateway",
"/home/user/app/tester",
"/home/user/app/lib"
]
for d in dirs:
assert os.path.isdir(d), f"D... | # test_final_state.py
import os
import pytest
def test_success_log_exists():
"""
Validates that the integration tester completed successfully,
meaning all concurrency, encoding, and ABI bugs were fixed.
"""
log_path = "/home/user/app/success.log"
assert os.path.exists(log_path), "success.log w... |
task_000000_0f9311e3 | task_000000_0f9311e3 | system_administration | Network | complex task (many commands spanning multiple phases: dependency installation, code writing, configuration, building, and testing) | bash and code (shell commands and writing/running scripts in Python, Perl, Ruby, etc.) | site reliability engineer monitoring uptime | Python | null | exact_text | text_only | legacy | null | no_verifiers | You are a Site Reliability Engineer responsible for taking over a legacy network segment. A proprietary local monitoring gateway is used to check the uptime of various backend microservices. Your goal is to write a Python script that provisions access for new staff, interacts with the legacy gateway to check service st... | The verification script will solely rely on checking the existence and contents of `/home/user/uptime_report.json`. It will also check that `/home/user/chroot/etc/shadow` contains properly formatted `$6$` hashes for the new hires.
Setup commands to initialize the environment:
```bash
# Create directory structure
mkdir... | [
"User account and group administration",
"Expect scripting for interactive automation",
"Connectivity diagnostics"
] | {"ablation": "no_verifiers", "anchor": null, "base_image": null, "command_complexity": "bash and code (shell commands and writing/running scripts in Python, Perl, Ruby, etc.)", "corpus_kind": "legacy", "description": "You are a Site Reliability Engineer responsible for taking over a legacy network segment. A proprietar... | #!/bin/bash
set -euo pipefail
mkdir -p /home/user/.setup_tmp
export PYTHONPATH=/home/user/.local/lib/python3/dist-packages:${PYTHONPATH:-}
apt-get update && apt-get install -y python3 python3-pip procps
pip3 install --target /home/user/.local/lib/python3/dist-packages pytest pexpect
# Create directory structure
mkdir -... | Bootstrap: docker
From: ubuntu:22.04
%post
export DEBIAN_FRONTEND=noninteractive
apt-get update && apt-get install -y python3 python3-pip procps
pip3 install pytest pexpect
# Create directory structure
mkdir -p /home/user/chroot/etc/
# Setup base chroot auth files
cat << 'EOF' > /home/user/chroot/etc/group
root:x:0:... | # test_initial_state.py
import os
import json
import socket
import pytest
def test_chroot_etc_files_exist():
base_dir = "/home/user/chroot/etc"
assert os.path.isdir(base_dir), f"Missing directory: {base_dir}"
for filename in ["group", "passwd", "shadow"]:
filepath = os.path.join(base_dir, filenam... | # test_final_state.py
import os
import json
import socket
import crypt
import pytest
def test_uptime_report():
report_path = "/home/user/uptime_report.json"
assert os.path.isfile(report_path), f"Missing report file: {report_path}"
with open(report_path, "r") as f:
try:
report = json.l... |
task_000000_12a3ab26 | task_000000_12a3ab26 | file_operations | File I/O | moderate task (several commands across setup, implementation, and verification) | bash-only (shell built-ins, coreutils, and standard CLI tools) | technical writer organizing documentation | Python | an archive extraction that overwrites files outside the target directory (zip slip) | exact_text | text_only | legacy | null | no_verifiers | You are a technical writer managing a docs-as-code repository. You have received an automated documentation bundle located at `/home/user/incoming_docs.zip`.
Due to an error in the upstream packaging script, this archive suffers from a "Zip Slip" vulnerability: some file paths contain dangerous directory traversal se... | The setup requires creating the vulnerable ZIP file and the INI configuration file before the agent begins.
**Setup Steps (to be executed by the test environment prior to agent invocation):**
```python
import os
import zipfile
# Ensure directories exist
os.makedirs('/home/user/docs_safe', exist_ok=True)
# Create the... | [
"Configuration file interpretation",
"Metadata-based file search",
"Large-scale text editing and macro application"
] | {"ablation": "no_verifiers", "anchor": "an archive extraction that overwrites files outside the target directory (zip slip)", "base_image": null, "command_complexity": "bash-only (shell built-ins, coreutils, and standard CLI tools)", "corpus_kind": "legacy", "description": "You are a technical writer managing a docs-as... | #!/bin/bash
set -euo pipefail
mkdir -p /home/user/.setup_tmp
export PYTHONPATH=/home/user/.local/lib/python3/dist-packages:${PYTHONPATH:-}
apt-get update && apt-get install -y python3 python3-pip unzip
python3 << 'EOF'
import os
import zipfile
# Ensure directories exist
os.makedirs('/home/user/docs_safe', exist_ok=Tru... | Bootstrap: docker
From: ubuntu:22.04
%post
export DEBIAN_FRONTEND=noninteractive
apt-get update && apt-get install -y python3 python3-pip unzip
pip3 install pytest
useradd -m -s /bin/bash user || true
python3 << 'EOF'
import os
import zipfile
# Ensure directories exist
os.makedirs('/home/user/docs_safe', exist_ok=T... | # test_initial_state.py
import os
import zipfile
import configparser
import pytest
def test_docs_safe_dir_exists():
"""Verify that the safe extraction directory exists."""
dir_path = '/home/user/docs_safe'
assert os.path.isdir(dir_path), f"Directory {dir_path} is missing. It should have been created durin... | # test_final_state.py
import os
import zipfile
import configparser
import pytest
DOCS_SAFE_DIR = '/home/user/docs_safe'
ZIP_PATH = '/home/user/incoming_docs.zip'
INI_PATH = '/home/user/metadata_rules.ini'
PUBLISHED_DOCS_TXT = '/home/user/published_docs.txt'
def parse_frontmatter(content):
"""Extracts the frontma... |
task_000000_1f43747e | task_000000_1f43747e | data_science | Systems | short task (a few shell commands focused on the core task) | bash-only (shell built-ins, coreutils, and standard CLI tools) | machine learning engineer preparing training data | any (model's choice) | null | exact_text | text_only | legacy | null | no_verifiers | You are a Machine Learning Engineer preparing training data and evaluating a baseline model. You need to process a dataset of hardware sensor logs, extract a text-based feature, reduce the dimensionality of your numerical features, and tune a simple classifier.
The dataset is located at: `/home/user/sensor_logs.csv`
... | The setup requires generating the dataset `/home/user/sensor_logs.csv` before the agent starts. No external libraries are needed for the setup script to ensure it runs cleanly on any baseline image.
Save the following Python script as `/tmp/setup_data.py` and run it during environment initialization:
```python
import... | [
"Model training and evaluation",
"Dimensionality reduction",
"Numerical accuracy testing",
"Tokenization and dataset preparation",
"Cross-validation and hyperparameter tuning"
] | {"ablation": "no_verifiers", "anchor": null, "base_image": null, "command_complexity": "bash-only (shell built-ins, coreutils, and standard CLI tools)", "corpus_kind": "legacy", "description": "You are a Machine Learning Engineer preparing training data and evaluating a baseline model. You need to process a dataset of ... | #!/bin/bash
set -euo pipefail
mkdir -p /home/user/.setup_tmp
export PYTHONPATH=/home/user/.local/lib/python3/dist-packages:${PYTHONPATH:-}
apt-get update && apt-get install -y python3 python3-pip
pip3 install --target /home/user/.local/lib/python3/dist-packages pytest pandas scikit-learn
cat << 'EOF' > /hom... | Bootstrap: docker
From: ubuntu:22.04
%post
export DEBIAN_FRONTEND=noninteractive
apt-get update && apt-get install -y python3 python3-pip
pip3 install pytest pandas scikit-learn
useradd -m -s /bin/bash user || true
cat << 'EOF' > /tmp/setup_data.py
import random
def setup_environment():
rand... | # test_initial_state.py
import os
def test_sensor_logs_dataset():
csv_path = '/home/user/sensor_logs.csv'
# Verify the file exists
assert os.path.exists(csv_path), f"The dataset file {csv_path} is missing."
assert os.path.isfile(csv_path), f"The path {csv_path} exists but is not a file."
# Verify... | # test_final_state.py
import os
import json
import math
import pytest
import tempfile
import random
import subprocess
import sys
def generate_pristine_dataset(filepath):
"""
Reconstructs the original dataset deterministically using the exact logic from the setup.
This ensures that even if the student modi... |
task_000000_20e21fd6 | task_000000_20e21fd6 | scientific_computing | Testing | short task (a few shell commands focused on the core task) | bash-only (shell built-ins, coreutils, and standard CLI tools) | researcher running simulations | Bash | a simulation that produces non-reproducible results due to floating-point reduction order | exact_text | text_only | legacy | null | no_verifiers | I am a bioinformatics researcher working on a project to estimate the binding affinities of various primer sequences using a Monte Carlo simulation.
I wrote a Bash pipeline to calculate a numerical integral over the GC-content of each primer sequence and output a score. To speed things up, I parallelized the workload... | The task requires fixing a race condition/ordering issue in a bash pipeline that affects floating point precision.
**Setup Instructions:**
The following script should be run to initialize the environment before the agent starts. This will create the primer list, the simulation script, the broken script, and the refer... | [
"Monte Carlo simulation",
"Primer design and sequence alignment",
"Reference dataset comparison",
"Numerical integration and differentiation",
"Parallel computing setup (MPI, OpenMP)"
] | {"ablation": "no_verifiers", "anchor": "a simulation that produces non-reproducible results due to floating-point reduction order", "base_image": null, "command_complexity": "bash-only (shell built-ins, coreutils, and standard CLI tools)", "corpus_kind": "legacy", "description": "I am a bioinformatics researcher workin... | #!/bin/bash
set -euo pipefail
mkdir -p /home/user/.setup_tmp
export PYTHONPATH=/home/user/.local/lib/python3/dist-packages:${PYTHONPATH:-}
apt-get update && apt-get install -y python3 python3-pip gawk findutils coreutils
mkdir -p /home/user
# 1. Generate primers.txt
cat << 'EOF' > /home/user/primers.txt... | Bootstrap: docker
From: ubuntu:22.04
%post
export DEBIAN_FRONTEND=noninteractive
apt-get update && apt-get install -y python3 python3-pip gawk findutils coreutils
pip3 install pytest
mkdir -p /home/user
# 1. Generate primers.txt
cat << 'EOF' > /home/user/primers.txt
1 ATGCGTA
2 GCATCGT
3 CGTA... | # test_initial_state.py
import os
import pytest
def test_primers_txt_exists_and_correct():
path = "/home/user/primers.txt"
assert os.path.isfile(path), f"File {path} is missing."
with open(path, "r") as f:
lines = f.read().strip().split("\n")
assert len(lines) == 50, f"Expected 50 lines in {pa... | # test_final_state.py
import os
import re
import pytest
import subprocess
def test_run_fixed_sh_exists_and_executable():
path = "/home/user/run_fixed.sh"
assert os.path.isfile(path), f"File {path} is missing. You must create it."
assert os.access(path, os.X_OK), f"File {path} is not executable. Make sure ... |
task_000000_23b202d9 | task_000000_23b202d9 | debugging | Algorithmic | short task (a few shell commands focused on the core task) | bash and code (shell commands and writing/running scripts in Python, Perl, Ruby, etc.) | operations engineer triaging incidents | Rust | null | exact_text | text_only | legacy | null | no_verifiers | "You are an Operations Engineer responding to an incident. A recent deployment of a custom Rust-base(...TRUNCATED) | "The verification of this task relies on automated scripts checking the build status, test suite, an(...TRUNCATED) | ["Regression test construction","Intermediate state tracing","Dependency conflict resolution","Forma(...TRUNCATED) | "{\"ablation\": \"no_verifiers\", \"anchor\": null, \"base_image\": null, \"command_complexity\": \"(...TRUNCATED) | "#!/bin/bash\nset -euo pipefail\nmkdir -p /home/user/.setup_tmp\nexport PYTHONPATH=/home/user/.local(...TRUNCATED) | "Bootstrap: docker\nFrom: ubuntu:22.04\n\n%post\nexport DEBIAN_FRONTEND=noninteractive\napt-get upda(...TRUNCATED) | "# test_initial_state.py\n\nimport os\n\ndef test_directories_exist():\n dirs = [\n \"/hom(...TRUNCATED) | "# test_final_state.py\n\nimport os\nimport json\nimport subprocess\nimport re\n\ndef test_cargo_bui(...TRUNCATED) |
task_000000_248ef650 | task_000000_248ef650 | data_querying | Data Comprehension | moderate task (several commands across setup, implementation, and verification) | bash and code (shell commands and writing/running scripts in Python, Perl, Ruby, etc.) | data engineer building ETL pipelines | any (model's choice) | null | exact_text | text_only | legacy | null | no_verifiers | "You are a data engineer responsible for analyzing user journey patterns on an e-commerce platform. (...TRUNCATED) | "The setup requires generating the SQLite database with specific planted graph patterns before the a(...TRUNCATED) | ["Query plan interpretation and optimization","Window functions and analytical aggregation","Query-t(...TRUNCATED) | "{\"ablation\": \"no_verifiers\", \"anchor\": null, \"base_image\": null, \"command_complexity\": \"(...TRUNCATED) | "#!/bin/bash\nset -euo pipefail\nmkdir -p /home/user/.setup_tmp\nexport PYTHONPATH=/home/user/.local(...TRUNCATED) | "Bootstrap: docker\nFrom: ubuntu:22.04\n\n%post\n export DEBIAN_FRONTEND=noninteractive\n apt-(...TRUNCATED) | "# test_initial_state.py\n\nimport os\nimport sqlite3\nimport pytest\n\nDB_PATH = '/home/user/ecomme(...TRUNCATED) | "# test_final_state.py\n\nimport json\nimport os\nimport sqlite3\nfrom collections import defaultdic(...TRUNCATED) |
task_000000_2c2b5253 | task_000000_2c2b5253 | software_engineering | Algorithmic | "complex task (many commands spanning multiple phases: dependency installation, code writing, config(...TRUNCATED) | "bash, code, and system services (shell commands, scripts, package installation, service configurati(...TRUNCATED) | engineer setting up a polyglot build system from scratch | Bash | null | exact_text | text_only | legacy | null | no_verifiers | "You are an engineer tasked with setting up a custom, polyglot build system and orchestration tool p(...TRUNCATED) | "Pre-task setup (Run as root during container build):\n```bash\napt-get update && apt-get install -y(...TRUNCATED) | ["REST/GraphQL API construction","Diff and patch processing","Schema migration","Cross-compilation a(...TRUNCATED) | "{\"ablation\": \"no_verifiers\", \"anchor\": null, \"base_image\": null, \"command_complexity\": \"(...TRUNCATED) | "#!/bin/bash\nset -euo pipefail\nmkdir -p /home/user/.setup_tmp\nexport PYTHONPATH=/home/user/.local(...TRUNCATED) | "Bootstrap: docker\nFrom: ubuntu:22.04\n\n%post\nexport DEBIAN_FRONTEND=noninteractive\napt-get upda(...TRUNCATED) | "# test_initial_state.py\nimport os\nimport shutil\nimport pytest\n\ndef test_commands_installed():\(...TRUNCATED) | "# test_final_state.py\nimport os\nimport json\nimport sqlite3\nimport subprocess\nimport urllib.req(...TRUNCATED) |
task_000000_34993bba | task_000000_34993bba | data_science | Algorithmic | moderate task (several commands across setup, implementation, and verification) | bash and code (shell commands and writing/running scripts in Python, Perl, Ruby, etc.) | data scientist cleaning datasets | Rust | null | exact_text | text_only | legacy | null | no_verifiers | "You are a data scientist tasked with cleaning and building a reproducible pipeline for manufacturin(...TRUNCATED) | "# Setup the environment and initial data\nmkdir -p /home/user/data\n\ncat << 'EOF' > /home/user/dat(...TRUNCATED) | ["Reproducible pipeline construction","Multi-source data joining","Missing value and outlier handlin(...TRUNCATED) | "{\"ablation\": \"no_verifiers\", \"anchor\": null, \"base_image\": null, \"command_complexity\": \"(...TRUNCATED) | "#!/bin/bash\nset -euo pipefail\nmkdir -p /home/user/.setup_tmp\nexport PYTHONPATH=/home/user/.local(...TRUNCATED) | "Bootstrap: docker\nFrom: ubuntu:22.04\n\n%post\n export DEBIAN_FRONTEND=noninteractive\n apt-(...TRUNCATED) | "# test_initial_state.py\n\nimport os\n\ndef test_metadata_csv_content():\n path = \"/home/user/d(...TRUNCATED) | "# test_final_state.py\n\nimport csv\nimport os\nimport subprocess\nimport pytest\n\ndef compute_exp(...TRUNCATED) |
task_000000_383442e8 | task_000000_383442e8 | data_processing | Manipulation | "intricate task (multi-stage workflow combining (a) environment setup or package configuration, (b) (...TRUNCATED) | bash and code (shell commands and writing/running scripts in Python, Perl, Ruby, etc.) | log analyst investigating patterns | any (model's choice) | a CSV pipeline that silently drops rows containing embedded newlines | exact_text | stripped_binary | rl_v2 | intricate | no_verifiers | "You are a Lead Data Engineer investigating a series of sophisticated attacks on our infrastructure.(...TRUNCATED) | "The task requires the agent to reverse-engineer the required input format for `/app/legacy_scorer`,(...TRUNCATED) | ["Unicode and multi-language text processing","Summary statistics and aggregation","Data sampling an(...TRUNCATED) | "{\"ablation\": \"no_verifiers\", \"anchor\": \"a CSV pipeline that silently drops rows containing e(...TRUNCATED) | "#!/bin/bash\nset -euo pipefail\nmkdir -p /home/user/.setup_tmp\nexport PYTHONPATH=/home/user/.local(...TRUNCATED) | "Bootstrap: docker\nFrom: ubuntu:22.04\n\n%files\n /gpfs/scrubbed/hamishiv/tmax-ablation-generati(...TRUNCATED) | "# test_initial_state.py\n\nimport os\n\ndef test_legacy_scorer_exists_and_executable():\n path =(...TRUNCATED) | "# test_final_state.py\n\nimport os\nimport csv\nimport json\nimport math\nfrom collections import d(...TRUNCATED) |
End of preview. Expand in Data Studio
TMax Ablation: No Verifiers (5K)
This dataset contains 5,000 selected initial-generation terminal tasks for the no_verifiers ablation. Verifier diversity is removed by forcing verifier_kind = exact_text. All remaining generation axes retain the original sampling configuration, with 3,333 legacy tasks and 1,667 rl_v2 tasks.
Tasks were generated with gemini/gemini-3.1-pro-preview using the direct Gemini API, with axes seed 42. This export contains the selected corpus, preserving the original task IDs, metadata and task file bytes.
Generation stage
These tasks passed the generation pipeline's build and initial-state checks. No rubric labeling, repair pass, or agent rollout verification has been applied. The generated final-state verifiers and reference answers may still contain defects. The truth field is the generator's reference answer; it is not an independently validated solution.
This arm requested 6,000 candidates across 41 completed chunks. The survivor target is exactly 5,000, selected to satisfy the corpus quotas. provenance.json records the configuration and generator source checksums; export-report.json records counts and artifact checksums.
Files and schema
data/train-00000-of-00001.parquet: thetrainsplit, one row per task.tasks.zip: task directories containing the originaltask.json,setup.sh,container.def,test_initial_state.py,test_final_state.py, and other task-local files included in the file manifest.task-ids.json: the 5,000 canonical task directory IDs.task-files.jsonl: archive paths, byte counts and SHA-256 checksums for every included task file.export-report.jsonandprovenance.json: export validation and generation provenance.
The canonical task_id matches the directory in tasks.zip; name preserves the original generated name. The Parquet columns include all sampled metadata: domain, skill_type, task_complexity, command_complexity, scenario, language, anchor, verifier_kind, fixture_kind, corpus_kind, base_image, ablation, and primitive_skills. They also contain description, truth, the complete original metadata as metadata_json, and the source text of setup, container_def, test_initial_state, and test_final_state.
metadata_json preserves original JSON types and values. Nullable complexity values remain null. Task-local fixture files are preserved in the ZIP. Binary container images, execution logs, and API completion caches are omitted. The cached validation base was used only during generation validation; saved task definitions are unchanged.
Loading
from datasets import load_dataset
tasks = load_dataset("TMaxxx/TMax-Ablation-No-Verifiers-5K", split="train")
from huggingface_hub import hf_hub_download
from zipfile import ZipFile
archive = hf_hub_download("TMaxxx/TMax-Ablation-No-Verifiers-5K", "tasks.zip", repo_type="dataset")
with ZipFile(archive) as z:
z.extractall("tmax-no_verifiers")
The task archive provides environment definitions and tests for further verification and repair. It does not include prebuilt container images or an agent solve pass.
- Downloads last month
- 3,332