Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability
Paper • 2610.08448 • Published • 79
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
passed: bool
archive_bytes: int64
archive_sha256: string
path_in_repo: string
modal_image_id: string
manifest: struct<release: string, base_archive_sha256: string, minimum_glibc: string, router_wheel_url: string (... 426 chars omitted)
child 0, release: string
child 1, base_archive_sha256: string
child 2, minimum_glibc: string
child 3, router_wheel_url: string
child 4, router_wheel_sha256: string
child 5, router_version: string
child 6, changed_python_files: list<item: string>
child 0, item: string
child 7, host_libraries: struct<libgcc_s.so.1: struct<source: string, sha256: string>, libstdc++.so.6: struct<source: string, (... 68 chars omitted)
child 0, libgcc_s.so.1: struct<source: string, sha256: string>
child 0, source: string
child 1, sha256: string
child 1, libstdc++.so.6: struct<source: string, sha256: string>
child 0, source: string
child 1, sha256: string
child 2, libz.so.1: struct<source: string, sha256: string>
child 0, source: string
child 1, sha256: string
child 8, extra_router_library_files: list<item: null>
child 0, item: null
child 9, notes: list<item: string>
child 0, item: string
child 10, files: list<item: struct<path: string, bytes: int64, sha256: string>>
child 0, item: struct<path: string, bytes: int64, sha256: string>
child 0, path: string
child 1, bytes: int64
child 2, sha256: string
toolkit_zip_sha256: string
qualification_report_sha256: string
toolkit_zip_bytes: int64
qualified_toolkit_commit: string
source_manifest_sha256: string
contents: string
to
{'archive_sha256': Value('string'), 'archive_bytes': Value('int64'), 'qualified_toolkit_commit': Value('string'), 'source_manifest_sha256': Value('string'), 'qualification_report_sha256': Value('string'), 'toolkit_zip_sha256': Value('string'), 'toolkit_zip_bytes': Value('int64'), 'contents': 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
passed: bool
archive_bytes: int64
archive_sha256: string
path_in_repo: string
modal_image_id: string
manifest: struct<release: string, base_archive_sha256: string, minimum_glibc: string, router_wheel_url: string (... 426 chars omitted)
child 0, release: string
child 1, base_archive_sha256: string
child 2, minimum_glibc: string
child 3, router_wheel_url: string
child 4, router_wheel_sha256: string
child 5, router_version: string
child 6, changed_python_files: list<item: string>
child 0, item: string
child 7, host_libraries: struct<libgcc_s.so.1: struct<source: string, sha256: string>, libstdc++.so.6: struct<source: string, (... 68 chars omitted)
child 0, libgcc_s.so.1: struct<source: string, sha256: string>
child 0, source: string
child 1, sha256: string
child 1, libstdc++.so.6: struct<source: string, sha256: string>
child 0, source: string
child 1, sha256: string
child 2, libz.so.1: struct<source: string, sha256: string>
child 0, source: string
child 1, sha256: string
child 8, extra_router_library_files: list<item: null>
child 0, item: null
child 9, notes: list<item: string>
child 0, item: string
child 10, files: list<item: struct<path: string, bytes: int64, sha256: string>>
child 0, item: struct<path: string, bytes: int64, sha256: string>
child 0, path: string
child 1, bytes: int64
child 2, sha256: string
toolkit_zip_sha256: string
qualification_report_sha256: string
toolkit_zip_bytes: int64
qualified_toolkit_commit: string
source_manifest_sha256: string
contents: string
to
{'archive_sha256': Value('string'), 'archive_bytes': Value('int64'), 'qualified_toolkit_commit': Value('string'), 'source_manifest_sha256': Value('string'), 'qualification_report_sha256': Value('string'), 'toolkit_zip_sha256': Value('string'), 'toolkit_zip_bytes': Value('int64'), 'contents': 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.
Cross-Tokenizer OPD B200 runtime
Portable runtime for arXiv 2610.08448v1, built from the authors' image and source.
The runtime archive is ctopd-b200-cu129-v4/runtime.tar.gz (16,515,065,867 bytes).
SHA-256: 084c321c0b7af1cd3a36d12a595c623567d4a37525aa0eecbc3933aa7f9b15c3.
Versions: Python 3.12.12, Torch 2.11.0+cu129, Transformers 5.6.2, SGLang 0.5.12.post1, FlashAttention 2.8.3, FlashInfer 0.6.11.post1.
Base image: docker.io/songmzhang/kdflow@sha256:bc49277f548992394dd9c583cded6d60d20ba73e0aa84f380e0fb7aac13691db.
The bundle materializes editable SGLang, includes distro 1.9.0 and standalone ninja 1.13.0.
Model weights, datasets, tokens and checkpoints are excluded.
Qualification and target
Six offline synthetic probes passed on one Modal B200: BF16, imports, packed FlashAttention forward/backward, FlashInfer, FSDP2/top-16 RKL, and patched SGLang generation/hidden states/logprobs/sleep-wake. The final suite took 112.550454101 seconds.
The maintained helper supports Ubuntu 22.04/glibc 2.35 and Ubuntu 24.04/glibc 2.39 on Linux x86_64, with g++ for JIT and an NVIDIA B200 driver compatible with CUDA 12.9. On glibc 2.35 it automatically applies the pinned 30,628,766-byte compat/glibc235-v2/overlay.tar.gz; on glibc >= 2.39 it retains the original v4 path. The base archive is unchanged. Allow about 60 GiB for the archive and installed tree together.
Tests use a random small model, world size/TP 1 and disabled SGLang CUDA graph capture. Company pods, production training, multi-GPU communication and paper results have not been validated.
An additional synthetic top_logprobs_num=4 fixture hit an error in pinned SGLang; that option is not used by the source repo's generate path and is not qualified here.
The Ubuntu 22.04/glibc 2.35 overlay passed all six B200 infrastructure probes in 80.143386344 seconds, plus real router registration and generation forwarding against a CPU stub. The original glibc 2.39 runtime passed a fresh CPU regression. Exact compatibility qualification and hashes are versioned separately. The matched router Python/native package is from one pinned same-version manylinux wheel; research code and Torch/CUDA are unchanged. A multiprocessing resource-tracker cleanup warning followed the GPU pass markers; exit status was zero.
Restore/check on a new company node
ensure_runtime.sh and ensure_runtime.SHA256SUMS are available at the repository root. Download both at the same immutable HF commit, check the checksum, and place them in /workspace/storage-shared/nlp/tungks/ctopd_all/. The helper is self-contained; it needs neither the toolkit ZIP nor GitHub access.
cd /workspace/storage-shared/nlp/tungks/ctopd_all
sha256sum -c ensure_runtime.SHA256SUMS
bash ensure_runtime.sh
Run inside a pod exposing the one B200 allocated to this check. It finds the archive in either ctopd-b200-cu129-v4/runtime.tar.gz or runtime.tar.gz, verifies its pinned size/SHA-256, and installs it into the node-local /opt/ctopd-b200 if absent. Allow at least 45 GiB free there. Existing identified runtimes are reused without rereading the full archive. Unknown or damaged installations are preserved for inspection rather than deleted/replaced. Set CTOPD_RUNTIME_PREFIX or --prefix to choose another explicit node-local directory if /opt is not writable.
Use --install-only before a GPU allocation (reports installed but unchecked), or --check-only to check without extraction. Default readiness checks pinned package versions, NVRTC headers, imports and a one-element CUDA memory roundtrip. It performs no model/data downloads, dependency builds, training or reduced training workload. It does not qualify full training or all kernels. Company-node execution remains unverified.
The maintained source and documentation are in sontungkieu/ctopd.
Choose the allocated GPU on a shared node
The helper preserves CUDA_VISIBLE_DEVICES; CUDA logical device 0 is not necessarily physical GPU 0. Use CUDA_VISIBLE_DEVICES=2 bash ensure_runtime.sh --prefix /opt/ctopd-b200-u22 --check-only, or the new convenience option bash ensure_runtime.sh --gpu 2 --prefix /opt/ctopd-b200-u22 --check-only. A GPU UUID is also accepted. --gpu explicitly overrides the inherited mask for both the probe and a Python command passed after --; without the flag, the mask is preserved. Selection does not allocate or reserve a GPU. Use only the GPU assigned to you, and retain the same visibility when launching training separately.
Download and install
Use the current ensure_runtime.sh above for both supported baselines. The historical toolkit installer/preflight below implements the original Ubuntu 24 runtime only and does not apply the Ubuntu 22 overlay.
This repository is public; no HF token is needed. download_ctopd.py uses 16 parallel HTTP byte ranges for the runtime itself, resumes partial ranges, and checks the complete archive and toolkit SHA-256. It requires Linux Python 3.9+ and only the standard library. Allow at least 35 GiB free during download/assembly, or about 60 GiB including installation.
Download download_ctopd.py and download_ctopd.SHA256SUMS from the same published commit, then on the company node:
set -euo pipefail
cd /workspace/storage-shared/nlp/tungks
mkdir -p ctopd_all
cd ctopd_all
# Place the two downloader files here first.
sha256sum -c download_ctopd.SHA256SUMS
python3 download_ctopd.py --workers 16 --directory "$PWD"
sha256sum -c SHA256SUMS toolkit.SHA256SUMS
unzip toolkit.zip -d toolkit
bash toolkit/company_preflight.sh
python3.12 toolkit/install_runtime.py ./runtime.tar.gz \
--sha256 084c321c0b7af1cd3a36d12a595c623567d4a37525aa0eecbc3933aa7f9b15c3 \
--target "$PWD/runtime"
Repeat the same Python download command with the same worker count after an interruption. Files are pinned to artifact revision 6058422a8319cbe03958bf5d52750293e4e0da65. The shared parent path is the previously supplied company location; it must already be mounted. The downloader and installation commands have not been executed on the company node.
Alternatively, with an existing HF CLI, set REPO=codemaivanngu/cross-tokenizer-opd-b200-runtime and REVISION=6058422a8319cbe03958bf5d52750293e4e0da65:
hf download "$REPO" --repo-type dataset --revision "$REVISION" \
--include 'ctopd-b200-cu129-v4/*' 'toolkit.zip' --local-dir ./ctopd-runtime
cd ctopd-runtime/ctopd-b200-cu129-v4
sha256sum -c SHA256SUMS
toolkit.zip contains the audited upstream source snapshot, installer, host preflight, offline probes, and qualification results. The source service did not expose a Git commit; the 84-file content snapshot is identified by source-manifest.json.
unzip ../toolkit.zip -d ../toolkit
bash ../toolkit/company_preflight.sh
python3.12 ../toolkit/install_runtime.py ./runtime.tar.gz \
--sha256 084c321c0b7af1cd3a36d12a595c623567d4a37525aa0eecbc3933aa7f9b15c3 \
--target /opt/ctopd-b200
SRC="$(cd ../toolkit/upstream && pwd)"
PYTHONPATH="$SRC" bash /opt/ctopd-b200/python-b200.sh -m kdflow.cli.train_kd_on_policy --help
The installer refuses an existing prefix. Package inventory retains author image wheel/editable provenance and is informational; install using the checked bundle. Original third-party software licenses remain applicable; this repository does not grant a new license for the bundled dependencies or upstream research code.
- Downloads last month
- 75