Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 81, in _split_generators
                  first_examples = list(islice(pipeline, self.NUM_EXAMPLES_FOR_FEATURES_INFERENCE))
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 32, in _get_pipeline_from_tar
                  fs: fsspec.AbstractFileSystem = fsspec.filesystem("memory")
                                                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 302, in filesystem
                  cls = get_filesystem_class(protocol)
                File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 239, in get_filesystem_class
                  raise ValueError(f"Protocol not known: {protocol}")
              ValueError: Protocol not known: memory
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 71, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

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.

Antidistillation Sampling — intermediate traces

Intermediate artifacts (teacher reasoning traces + run metadata) from antidistillation sampling experiments with an openai/gpt-oss-120b teacher. Shared for experiment reproduction.

Each experiment is a gzipped tarball with its own README.md inside describing the setup, the file inventory, and the measured accuracies.

archive size run max gen length what's in it
exp7.tar.gz 54 MB 4k-context run, 2026-05-13 → 05-21 4096 holdout traces (exp9's holdout set lives here) + 8 small probe runs + 4 full-scale sets (all degenerate — see its README)
exp8.tar.gz 16 MB 8k-context run, 2026-05-14 → 05-20 8192 holdout traces + one lam=0.04 training set (degenerate — see its README)
exp9.tar.gz 34 MB 2k-context run, 2026-05-21 → 05-24 2048 full lam sweep: 5 training-trace sets (lam 0.01→0.02) + 7 teacher evals (lam 0.0→0.3)

Start with exp9 — it is the most recent and the only one with a usable lam sweep. It has no holdout set of its own; the matching one is exp7/traces/holdout.parquet (their gradients are byte-identical).

exp7 and exp8 are included for the context-length comparison and for the record, but their full-scale training traces all collapse to ~0.001 accuracy with degenerate token repetition. In exp7's case this happens even at lam=0.0001, where ADS is effectively off, so it is more likely a generation bug than the defense working. Each package's README spells this out per file.

Shared setup

Role Model
Teacher (protected) openai/gpt-oss-120b, MXFP4
Proxy student (ADS gradients) openai/gpt-oss-20b
Downstream student (distillation target) meta-llama/Llama-3.2-1B

Datasets: Big-Math (training + holdout), GSM8K (teacher eval). Sampling: tau=0.6, eps=1e-2, seed=42.

Usage

tar xzf exp9.tar.gz
python -c "
import pandas as pd
d = pd.read_parquet('exp9/traces/tau0.6_lam0.01_eps1.00e-02.parquet')
print(d.shape, list(d.columns))
print('teacher accuracy:', d.is_raw_correct.mean())
"

All parquet files share the schema problem, solution, trace, is_raw_correct, where trace is the teacher's raw output in gpt-oss harmony format. Every trace file has a matching .yaml config snapshot, and metadata/trace_registry.jsonl records the full hyperparameter set (lam, tau, eps, seed, split, models, sample count) for each one.

Proxy-student gradients

The Stage-2 gradients used by ADS, computed with the openai/gpt-oss-20b proxy student. Stored at the repo root rather than inside the tarballs, because two of the three runs share one file:

file size used by computed from
student_grads_exp7_exp9.pt 3.6 GB exp7 and exp9 (byte-identical) exp7/traces/holdout.parquet
student_grads_exp8.pt 3.6 GB exp8 exp8/traces/holdout.parquet

Do not mix them up — pairing a run with the wrong gradient file will silently produce wrong ADS behavior.

These are torch.save pickles, so loading them executes arbitrary code; load only from a source you trust:

grads = torch.load("student_grads_exp7_exp9.pt", map_location="cpu")

You need them only to re-run ADS sampling itself. Reproducing the distillation or the accuracies reported in each package does not require them — the traces are already generated.

Not included

The sharded HuggingFace datasets directories sitting next to each parquet (same content, larger), and distilled student checkpoints (none were produced for these three runs).

Verifying downloads

a79cb78c046fb842453553dd02a69a1326ba48dd4ebec0b2016cd1b589e2ec41  exp7.tar.gz
422a5e3a1a8f301c15b62fef7fbe51a8afedf70937faffa4898a308e9e71e9c3  exp8.tar.gz
d82e1ab3a53075ec37944936ea66badc5307a7549a4cb57da5099b72af0256bc  exp9.tar.gz
09e00e592256745714197ef5c9e84e51ab297e22123679db35a2b73228a2329c  student_grads_exp7_exp9.pt
4e815f25e6c8fc1ad08d1c97e61ac3634e28e53df880c5362b7b7ca7c423b131  student_grads_exp8.pt

Or verify everything at once with the included SHA256SUMS:

sha256sum -c SHA256SUMS

Provenance and licensing

  • Problems and reference answers derive from Big-Math (CC-BY-4.0) and GSM8K (MIT) — attribution required for redistribution.
  • Traces are outputs of openai/gpt-oss-120b (Apache-2.0); the model license places no restriction on generated outputs.
  • Built with Llama: the surrounding pipeline distills into meta-llama/Llama-3.2-1B.

The license: other tag reflects this mix — the underlying licenses above govern use.

Downloads last month
49