Breaking the Tokenizer Barrier: On-Policy Distillation across Model Families
Paper • 2606.09456 • Published
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
assets: list<item: struct<file_count: int64, repo_id: string, revision: string, role: string, total_bytes: i (... 6 chars omitted)
child 0, item: struct<file_count: int64, repo_id: string, revision: string, role: string, total_bytes: int64>
child 0, file_count: int64
child 1, repo_id: string
child 2, revision: string
child 3, role: string
child 4, total_bytes: int64
elapsed_seconds: double
gpu_used: bool
hashed_files_verified: int64
manifest: string
platform: string
profile: string
python: string
recorded_unix: int64
schema_version: int64
stage: string
status: string
total_bytes: int64
weights_hashed: bool
upstream_repo: string
upstream_commit: string
to
{'assets': List({'file_count': Value('int64'), 'files': List({'bytes': Value('int64'), 'path': Value('string'), 'sha256': Value('string')}), 'local_dir': Value('string'), 'repo_id': Value('string'), 'revision': Value('string'), 'role': Value('string'), 'total_bytes': Value('int64'), 'upstream_repo_id': Value('string'), 'upstream_revision': Value('string'), 'weights_hashed': Value('bool')}), 'schema_version': Value('int64'), 'total_bytes': Value('int64'), 'upstream_commit': Value('string'), 'upstream_repo': Value('string'), 'weights_hashed': Value('bool')}
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
assets: list<item: struct<file_count: int64, repo_id: string, revision: string, role: string, total_bytes: i (... 6 chars omitted)
child 0, item: struct<file_count: int64, repo_id: string, revision: string, role: string, total_bytes: int64>
child 0, file_count: int64
child 1, repo_id: string
child 2, revision: string
child 3, role: string
child 4, total_bytes: int64
elapsed_seconds: double
gpu_used: bool
hashed_files_verified: int64
manifest: string
platform: string
profile: string
python: string
recorded_unix: int64
schema_version: int64
stage: string
status: string
total_bytes: int64
weights_hashed: bool
upstream_repo: string
upstream_commit: string
to
{'assets': List({'file_count': Value('int64'), 'files': List({'bytes': Value('int64'), 'path': Value('string'), 'sha256': Value('string')}), 'local_dir': Value('string'), 'repo_id': Value('string'), 'revision': Value('string'), 'role': Value('string'), 'total_bytes': Value('int64'), 'upstream_repo_id': Value('string'), 'upstream_revision': Value('string'), 'weights_hashed': Value('bool')}), 'schema_version': Value('int64'), 'total_bytes': Value('int64'), 'upstream_commit': Value('string'), 'upstream_repo': Value('string'), 'weights_hashed': Value('bool')}
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.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
DPCA portable B200 runtime
Generated by the Modal packaging step. Do not edit files here by hand: every file
is covered by SHA256SUMS and the Hub commit is the release identifier.
Paper under qualification: arXiv:2606.09456, teacher Qwen/Qwen3-8B -> student
Llama-3.1-8B-SFT (Appendix A, Table 4). Nothing in this payload is evidence
about that paper's results.
| file | role |
|---|---|
runtime.tar.gz |
the runtime itself (venv + base interpreter + pinned source) |
models.tar |
the paper pair (~32 GB), extracted into shared storage by stage_models.sh |
stage_models.sh |
verify + extract the weights into $DPCA_CACHE_ROOT, then re-verify against the manifest |
python-b200-dpca.sh |
entry point; sets PYTHONPATH, LD_LIBRARY_PATH, cache roots |
install.sh |
verify, unpack the runtime into /, run the smoke gate |
smoke.py |
first gate: does this runtime see a Blackwell GPU |
gates.py |
full gate code; identical file runs on Modal and on the node |
pins.py |
every pinned version/commit this payload was built from |
stage_assets.py |
stages the paper pair into shared storage, pinned by revision |
assets-manifest.json |
per-file repo/revision/size/sha256 manifest written by stage_assets.py stage; stage_models.sh copies it into the cache and re-verifies every weight against it |
assets-qualification.json |
the assets gate receipt: what the gate observed on Modal |
manifest.json |
what was packaged, from which image, with which profile |
packages.json |
full pip-style inventory of the venv |
SHA256SUMS |
checksums for every payload file |
Order of operations on the node
- read-only preflight (
nvidia-smi,python3 --version, proxy reachability); - download this payload through the corporate proxy; verify
SHA256SUMS; bash install.sh;DPCA_CACHE_ROOT=/workspace/storage-shared/... bash stage_models.sh(weights land in shared storage, never in/);bash python-b200-dpca.sh gates.py cpu;bash python-b200-dpca.sh gates.py cuda;bash python-b200-dpca.sh gates.py dpca_gpu;DPCA_CACHE_ROOT=... bash python-b200-dpca.sh gates.py assets.
Steps 1-8 are engineering qualification. They say nothing about scientific results, and the canary receipt says the same about itself.
- Downloads last month
- 257