Datasets:
The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
task: string
test_stdout: string
test_results: struct<results: struct<tool: struct<name: string, version: string>, summary: struct<tests: int64, pa (... 321 chars omitted)
child 0, results: struct<tool: struct<name: string, version: string>, summary: struct<tests: int64, passed: int64, fai (... 304 chars omitted)
child 0, tool: struct<name: string, version: string>
child 0, name: string
child 1, version: string
child 1, summary: struct<tests: int64, passed: int64, failed: int64, skipped: int64, pending: int64, other: int64, sta (... 25 chars omitted)
child 0, tests: int64
child 1, passed: int64
child 2, failed: int64
child 3, skipped: int64
child 4, pending: int64
child 5, other: int64
child 6, start: double
child 7, stop: double
child 2, tests: list<item: struct<name: string, status: string, duration: double, start: double, stop: double, retri (... 108 chars omitted)
child 0, item: struct<name: string, status: string, duration: double, start: double, stop: double, retries: int64, (... 96 chars omitted)
child 0, name: string
child 1, status: string
child 2, duration: double
child 3, start: double
child 4, stop: double
child 5, retries: int64
child 6, file_path: string
child 7, raw_status: string
child 8, trace: string
child 9, message: string
child 10, tags: list<item: string>
child 0, item: string
model: string
n_tasks: int64
seed: int64
benchmark: string
pass_at_1: double
tasks: list<item: struct<task: string, reward: double, exception: string>>
child 0, item: struct<task: string, reward: double, exception: string>
child 0, task: string
child 1, reward: double
child 2, exception: string
to
{'model': Value('string'), 'benchmark': Value('string'), 'seed': Value('int64'), 'n_tasks': Value('int64'), 'pass_at_1': Value('float64'), 'tasks': List({'task': Value('string'), 'reward': Value('float64'), 'exception': Value('string')})}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 483, 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 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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 343, 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 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
task: string
test_stdout: string
test_results: struct<results: struct<tool: struct<name: string, version: string>, summary: struct<tests: int64, pa (... 321 chars omitted)
child 0, results: struct<tool: struct<name: string, version: string>, summary: struct<tests: int64, passed: int64, fai (... 304 chars omitted)
child 0, tool: struct<name: string, version: string>
child 0, name: string
child 1, version: string
child 1, summary: struct<tests: int64, passed: int64, failed: int64, skipped: int64, pending: int64, other: int64, sta (... 25 chars omitted)
child 0, tests: int64
child 1, passed: int64
child 2, failed: int64
child 3, skipped: int64
child 4, pending: int64
child 5, other: int64
child 6, start: double
child 7, stop: double
child 2, tests: list<item: struct<name: string, status: string, duration: double, start: double, stop: double, retri (... 108 chars omitted)
child 0, item: struct<name: string, status: string, duration: double, start: double, stop: double, retries: int64, (... 96 chars omitted)
child 0, name: string
child 1, status: string
child 2, duration: double
child 3, start: double
child 4, stop: double
child 5, retries: int64
child 6, file_path: string
child 7, raw_status: string
child 8, trace: string
child 9, message: string
child 10, tags: list<item: string>
child 0, item: string
model: string
n_tasks: int64
seed: int64
benchmark: string
pass_at_1: double
tasks: list<item: struct<task: string, reward: double, exception: string>>
child 0, item: struct<task: string, reward: double, exception: string>
child 0, task: string
child 1, reward: double
child 2, exception: string
to
{'model': Value('string'), 'benchmark': Value('string'), 'seed': Value('int64'), 'n_tasks': Value('int64'), 'pass_at_1': Value('float64'), 'tasks': List({'task': Value('string'), 'reward': Value('float64'), 'exception': Value('string')})}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
MiRL Terminal-Bench 2.1 Trajectories
Trajectories for paper "MiRL: Error-Targeted Reinforcement Learning for Correcting Terminal Agent Misbehaviors".
We release the trajectories of an evaluation on Terminal-Bench 2.1 (89 tasks, 16 seeds per model) for the base model Qwen3.5-4B, the main baseline TMax-4B, and our model MiRL-Qwen3.5-4B.
| Model | Directory | Seeds | pass@1 (mean ± SE over seeds) |
|---|---|---|---|
| MiRL-Qwen3.5-4B | traj_MiRL/ |
16 | 21.1 ± 0.5 |
| TMax-4B | traj_TMax-4B/ |
16 | 20.2 ± 0.7 |
| Qwen3.5-4B (base) | traj_base/ |
16 | 14.2 ± 0.6 |
Unbiased pass@k (%) over the 16 seeds, averaged over the 89 tasks:
| Model | pass@1 | pass@2 | pass@4 | pass@8 | pass@16 |
|---|---|---|---|---|---|
| MiRL-Qwen3.5-4B | 21.1 | 25.9 | 30.3 | 34.7 | 39.3 |
| TMax-4B | 20.2 | 24.3 | 27.7 | 30.7 | 33.7 |
| Qwen3.5-4B (base) | 14.2 | 19.3 | 24.7 | 29.9 | 34.8 |
Layout
Each model directory has seed_00/ … seed_15/. Each seed directory has three files; in all of them the 89 tasks are
sorted by name, one task per line or entry:
trajectories.jsonl:task,messages(the full agent run: system prompt, task, and each step's reasoning, bash command and command output).tests.jsonl:task,test_stdout(output of the task's tests after the run),test_results(per-test results).scores.json:model,benchmark,seed,n_tasks,pass_at_1, and per taskrewardandexception.
Host-specific information (hostnames, IP addresses, paths) and secrets generated inside the sandboxes were replaced
with placeholders such as sandbox and <redacted>.
Code: https://github.com/runchu-tian/MiRL
Citation
@misc{tian2026mirlerrortargetedreinforcementlearning,
title={MiRL: Error-Targeted Reinforcement Learning for Correcting Terminal Agent Misbehaviors},
author={Runchu Tian and Zaid Khan and Zun Wang and Justin Chih-Yao Chen and Mohit Bansal},
year={2026},
eprint={2610.35697},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2610.35697},
}
- Downloads last month
- 463