Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillation
Paper • 2608.19098 • Published • 23
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: CastError
Message: Couldn't cast
input_prompt: string
output_response: string
score: double
stop_reason: string
exception_type: null
error_treatment: null
env_class: string
env_extras: struct<data_source: string, reward_model: struct<ground_truth: string>, extra_info: struct<source_id (... 28 chars omitted)
child 0, data_source: string
child 1, reward_model: struct<ground_truth: string>
child 0, ground_truth: string
child 2, extra_info: struct<source_id: string>
child 0, source_id: string
child 3, max_turns: int64
data_source: string
eval/aime_2024/pass_at_1: double
eval/aime_2024/avg_score: double
eval/all/pass_at_1: double
eval/all/avg_score: double
to
{'eval/aime_2024/avg_score': Value('float64'), 'eval/aime_2024/pass_at_1': Value('float64'), 'eval/all/avg_score': Value('float64'), 'eval/all/pass_at_1': Value('float64')}
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 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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
input_prompt: string
output_response: string
score: double
stop_reason: string
exception_type: null
error_treatment: null
env_class: string
env_extras: struct<data_source: string, reward_model: struct<ground_truth: string>, extra_info: struct<source_id (... 28 chars omitted)
child 0, data_source: string
child 1, reward_model: struct<ground_truth: string>
child 0, ground_truth: string
child 2, extra_info: struct<source_id: string>
child 0, source_id: string
child 3, max_turns: int64
data_source: string
eval/aime_2024/pass_at_1: double
eval/aime_2024/avg_score: double
eval/all/pass_at_1: double
eval/all/avg_score: double
to
{'eval/aime_2024/avg_score': Value('float64'), 'eval/aime_2024/pass_at_1': Value('float64'), 'eval/all/avg_score': Value('float64'), 'eval/all/pass_at_1': Value('float64')}
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.
MarinSkyRL native Open-MOPD trace archive
This dataset preserves the retained student trajectories and inline AIME evaluation outputs from the native MarinSkyRL Open-MOPD experiment. The selected step-32 checkpoint is the result of three linked job roots. Files retain the original schema-v3 archive names and directory structure. Source object-store files were copied on 2026-09-18 without rewriting their contents.
| Folder | Original run | Coverage | Relation to step-32 policy |
|---|---|---|---|
trajectories/original/ |
open-mopd-native-20260916c |
Student train steps 1–23 | Steps 1–4 are ancestors; 5–23 belong to the separately continued slower branch. |
trajectories/fast-gate/ |
open-mopd-native-fast-gate-20260917a |
Student train steps 5–6 | Ancestors seeded from original step 4. |
trajectories/fast-full/ |
open-mopd-native-fast-full-20260917a |
Student train steps 7–35; two archives at step 15 after retry | Steps 7–32 are ancestors; 33–35 are later trace output. The durable job checkpoint reached step 34. |
inline-evals/original/ |
Original run | AIME24 inline output at even checkpoints 2–8 | Steps 2 and 4 are on the selected lineage; 6 and 8 are on the slow branch. |
inline-evals/fast-gate/ |
Fast-gate run | AIME24 inline output at step 6 | Validation from the step-4-seeded gate. |
inline-evals/fast-full/ |
Fast-full run | AIME24 inline output at even checkpoints 8–34 | Validation traces, not teacher-scoring traces. |
Each trajectories/*/schema_v3/archives/phase=train/step=…/<sha256>.zip is a retention archive. Open manifest.json inside a zip to see its member records. Each records/*.json.gz contains one JSON object with the prompt, student response and token IDs, step, sample provenance, domain, and teacher_route. The zip filename is the SHA-256 of the zip bytes; verify it with shasum -a 256. Retention rules mean these are retained samples, not every generated rollout in every batch. See the sibling retention ledgers for the source retention decisions.
Teacher evidence boundary
The experiment used three pinned local inference teachers, one for each domain. The saved student record identifies which teacher route was selected. The original retention schema did not store the teacher's per-token log-probability arrays or an independent teacher service request/response trace. Therefore this dataset is not a complete teacher-and-student scoring transcript, and those missing historical teacher scores cannot be recovered exactly from this archive. Recompute teacher scores from the pinned teacher checkpoints and source code if you need a new analysis; that will be a new measurement, not the original recorded trace. The companion local artifacts/open-mopd-repro/launch-records/ bundle contains the model revisions, dataset hashes, source commits, and complete launch commands; it is intended for later publication as an artifacts repository and is not part of this trace dataset.
The retained prompts and responses come from the authors' public Open-MOPD data and inline AIME validation. This archive is research evidence, not a new training set or a paper-comparable benchmark result. The separate Open-MOPD paper describes the underlying method.
- Downloads last month
- 29