PIArena: A Platform for Prompt Injection Evaluation
Paper • 2604.08499 • Published • 1
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: TypeError
Message: Couldn't cast array of type
struct<target_inst: string, context: string, injected_task: string, target_task_answer: string, injected_task_answer: string, category: string, orig_id: string, orig_language: string, orig_scenario_type: string, orig_difficulty: string, orig_split_group: string, injected_context: string>
to
{'target_inst': Value('string'), 'context': Value('string'), 'injected_task': Value('string'), 'target_task_answer': Value('string'), 'injected_task_answer': Value('string'), 'category': Value('string'), 'orig_id': Value('string'), 'orig_language': Value('string'), 'orig_scenario_type': Value('string'), 'orig_difficulty': Value('string'), 'orig_split_group': Value('string'), 'defense_result': {'response': Value('string')}, 'utility': Value('bool'), 'asr': Value('bool')}
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2315, in cast_table_to_schema
cast_array_to_feature(
~~~~~~~~~~~~~~~~~~~~~^
table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
feature,
^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
~~~~^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2161, in cast_array_to_feature
raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
TypeError: Couldn't cast array of type
struct<target_inst: string, context: string, injected_task: string, target_task_answer: string, injected_task_answer: string, category: string, orig_id: string, orig_language: string, orig_scenario_type: string, orig_difficulty: string, orig_split_group: string, injected_context: string>
to
{'target_inst': Value('string'), 'context': Value('string'), 'injected_task': Value('string'), 'target_task_answer': Value('string'), 'injected_task_answer': Value('string'), 'category': Value('string'), 'orig_id': Value('string'), 'orig_language': Value('string'), 'orig_scenario_type': Value('string'), 'orig_difficulty': Value('string'), 'orig_split_group': Value('string'), 'defense_result': {'response': Value('string')}, 'utility': Value('bool'), 'asr': Value('bool')}Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
PIArena × networkadminquestion — results
Prompt-injection robustness evaluation of PIArena (ACL 2026, https://arxiv.org/abs/2604.08499) on banhcarrot/networkadminquestion converted to PIArena schema.
Setup
- Backend LLM:
Qwen/Qwen3-4B-Instruct-2507(local, greedy, default PIArena sampling) - Judge: same model via PIArena
llm_judge(Utility and ASR; judge sees task + response, not the gold answer) - Defense:
none - Data: 144 rows of the test split of
banhcarrot/networkadminquestionconverted to PIArena's 6-column schema — banhcarrot/networkadminquestion-piarena - Attack:
directinjection at the end of the user question; injected task is a Vietnamese malicious instruction rotating over 4 categories (phishing, content promotion, access denial, infrastructure failure) - Seed 42, single run.
Results (n = 144)
| run | attack | Utility | ASR |
|---|---|---|---|
| netadmin-baseline | none | 0.806 | 0.000 |
| netadmin-direct | direct | 0.840 | 0.028 (4/144) |
Interpretation notes
- ASR 2.8% under direct injection is a strong result: Qwen3-4B followed its system-prompt JSON discipline and ignored the injected Vietnamese instruction in 140/144 cases. In the smoke run (5 rows) a successful injection visibly leaked the phishing advice into an otherwise-correct tool-call JSON.
- Utility 0.806 at baseline reflects the LLM-judge's strictness on "exactly one JSON, canonical reason strings" — the judge does not see the gold
target_json(PIArena'sllm_judgedesign), so Utility measures formal task completion, not exact-match accuracy. - Utility slightly higher under attack (0.840 vs 0.806) is judge noise at this scale (±1 sample ≈ 0.007); not a real improvement.
Files
results/netadmin-baseline/…— baseline evaluation + attack cacheresults/netadmin-direct/…— direct-attack evaluation + attack cache- smoke-run artifacts (5 rows) from the pipeline validation are also included
Converted dataset: banhcarrot/networkadminquestion-piarena · Trackio dashboard: banhcarrot/piarena-networkadmin-trackio
- Downloads last month
- 63