Datasets:
task large_stringlengths 7 25 | messages listlengths 10 62 | reward float64 0 1 | pct_test_cases_passed float64 0 100 | n_test_cases_passed int64 0 5 | n_test_cases_total int64 1 7 | test_cases listlengths 1 7 | executed_commands listlengths 5 71 |
|---|---|---|---|---|---|---|---|
break-filter-js-from-html | [
{
"role": "system",
"content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON... | 0 | 0 | 0 | 1 | [
{
"name": "test_outputs.py::test_out_html_bypasses_filter",
"status": "failed"
}
] | [
"cat /app/filter.py",
"cat /app/test_outputs.py",
"cat /tests/filter.py 2>/dev/null || echo \"Not found\"",
"echo '<img src=\"javascript:alert(1)\">' > /app/out.html",
"python /app/filter.py /app/out.html",
"cat /app/out.html",
"echo '<svg onload=alert(1)>' > /app/out.html",
"python /app/filter.py /ap... |
build-pov-ray | [
{
"role": "system",
"content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON... | 0 | 0 | 0 | 3 | [
{
"name": "test_outputs.py::test_illum1_render_and_verify",
"status": "failed"
},
{
"name": "test_outputs.py::test_povray_version",
"status": "failed"
},
{
"name": "test_outputs.py::test_povray_built_from_correct_source",
"status": "failed"
}
] | [
"ls -la /app/",
"ls -la /app/deps/",
"which wget curl gnupg",
"which tar gzip",
"wget http://apnic.net.au/~chrylitz/povray/2.2/zip/povray-2.2.zip -O /app/povray-2.2.zip",
"rm /app/povray-2.2.zip",
"wget https://github.com/pei职权/povray/raw/master/povray-2.20.tar.gz -O /app/povray-2.20.tar.gz 2>&1",
"wg... |
cancel-async-tasks | [
{
"role": "system",
"content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON... | 0 | 16.67 | 1 | 6 | [
{
"name": "test_outputs.py::test_run_py_file_exists",
"status": "passed"
},
{
"name": "test_outputs.py::test_tasks_run_concurrently",
"status": "failed"
},
{
"name": "test_outputs.py::test_tasks_obey_max_concurrent",
"status": "failed"
},
{
"name": "test_outputs.py::test_task... | [
"mkdir -p /app && python3 --version",
"cat > /app/run.py << 'EOF'",
"cat > /app/run.py << 'EOF'",
"cat /app/run.py",
"python3 -c \"from run import run_tasks; print('Import OK')\"",
"python3 << 'EOF'",
"cat > /app/run.py << 'EOF'",
"python3 << 'EOF'",
"cat > /app/run.py << 'EOF'",
"python3 << 'EOF'... |
chess-best-move | [
{
"role": "system",
"content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON... | 0 | 0 | 0 | 1 | [
{
"name": "test_outputs.py::test_move_correct",
"status": "failed"
}
] | [
"ls -la",
"which python3 && python3 --version",
"python3 -c \"from PIL import Image; print('Pillow available')\"",
"python3 -c \"from PIL import Image; print('Pillow available')\"",
"python3 << 'EOF'",
"python3 << 'EOF'",
"python3 << 'EOF'",
"python3 << 'EOF'",
"python3 << 'EOF'",
"python3 << 'EOF... |
circuit-fibsqrt | [{"role":"system","content":"You are an AI assistant tasked with solving command-line tasks in a Lin(...TRUNCATED) | 0 | 66.67 | 2 | 3 | [{"name":"test_outputs.py::test_gates_file_exists","status":"passed"},{"name":"test_outputs.py::test(...TRUNCATED) | ["cat /app/sim.c","echo 'out0 = 10' | /app/sim 20","gcc -o /app/sim /app/sim.c","echo 'out0 = 1","ca(...TRUNCATED) |
constraints-scheduling | [{"role":"system","content":"You are an AI assistant tasked with solving command-line tasks in a Lin(...TRUNCATED) | 0 | 0 | 0 | 3 | [{"name":"test_outputs.py::test_structure_event_title_attendees_duration_window","status":"failed"},(...TRUNCATED) | ["cat /app/alice_calendar.ics","cat /app/bob_calendar.ics","cat /app/carol_calendar.ics","cat > /app(...TRUNCATED) |
crack-7z-hash | [{"role":"system","content":"You are an AI assistant tasked with solving command-line tasks in a Lin(...TRUNCATED) | 0 | 0 | 0 | 2 | [{"name":"test_outputs.py::test_solution_file","status":"failed"},{"name":"test_outputs.py::test_sol(...TRUNCATED) | ["ls -la","unzip -l secrets.7z","7z l secrets.7z","which p7zip ; find /usr -name '*.7z*' 2>/dev/null(...TRUNCATED) |
custom-memory-heap-crash | [{"role":"system","content":"You are an AI assistant tasked with solving command-line tasks in a Lin(...TRUNCATED) | 0 | 0 | 0 | 6 | [{"name":"test_outputs.py::test_protected_files_not_modified","status":"failed"},{"name":"test_outpu(...TRUNCATED) | ["ls -la /app/","cat /app/user.cpp","cat /app/main.cpp","cat /app/user.h","g++ -std=c++17 -o /app/re(...TRUNCATED) |
db-wal-recovery | [{"role":"system","content":"You are an AI assistant tasked with solving command-line tasks in a Lin(...TRUNCATED) | 0 | 0 | 0 | 7 | [{"name":"test_outputs.py::test_recovered_json_exists","status":"failed"},{"name":"test_outputs.py::(...TRUNCATED) | ["ls -la","ls -la main.db/","file main.db","ls -la main.db","python3 -c \"import sqlite3; conn = sql(...TRUNCATED) |
dna-assembly | [{"role":"system","content":"You are an AI assistant tasked with solving command-line tasks in a Lin(...TRUNCATED) | 0 | 0 | 0 | 1 | [
{
"name": "test_outputs.py::test_primers",
"status": "failed"
}
] | ["cat sequences.fasta","cat -A sequences.fasta | head -100","which primer3 && primer3 --version","ap(...TRUNCATED) |
End of preview. Expand in Data Studio
Terminal-Bench 2.0 eval — Qwen3.5-4B base (no warm-up)
Per-sample results of Qwen3.5-4B base (no warm-up) evaluated as an actor (terminus-2 agent prompt) on Terminal-Bench 2.0 (Harbor, Apptainer+fakeroot). One row per task; the hidden pytest oracle runs in the container after the agent acts.
- tasks: 31 · fully resolved (reward=1): 1 · mean test-cases passed: 26.2%
Columns
| column | type | meaning |
|---|---|---|
task |
str | task name |
messages |
list[{role, content}] | the [system, user, assistant, ...] conversation |
reward |
float | task reward (1.0 iff ALL test cases pass) |
pct_test_cases_passed |
float | partial credit, e.g. 60.0 |
n_test_cases_passed / n_test_cases_total |
int | e.g. 3 / 5 |
test_cases |
list[{name, status}] | per-pytest-case pass/fail |
executed_commands |
list[str] | commands the model ran (best-effort, from the terminal echo) |
Fidelity: harbor stores only the model's analysis/plan as the assistant message (not the raw
{commands:[...]} JSON); the executed command appears as the terminal echo in the next user/observation
turn and is also extracted into executed_commands.
Model: violetxi/qwen35-4b-terminal-wm-base ·
companion eval repos: base / wm-nextobs / wm-summary.
- Downloads last month
- 52