The dataset viewer is not available for this subset.
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 66, 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.
AI Plays Tag — Design C checkpoints
Complete training artifacts for the pre-registered reward shaping × opponent diversity factorial study in AI-Plays-Tag: 195 PPO zoo-self-play training runs (plus three 10M-step pilots and the 4 pre-registered anchor runs), 5M steps each, in a 2D tag environment.
Headline result (confirmatory, pre-registered): reward shaping and zoo
opponent mixing are complements — β_RA = +2.05 log-odds
[+0.72, +3.45], P(>0) = 0.999 — and dense shaping trades a +30pp mean
improvement for a 3–4× between-seed "seed lottery". Full study record:
experiments/design_c_results.md
(pre-registration chain, adversarial-review errata, mechanistic
follow-up). Evaluation CSVs (gauntlet, anchor panel, MCMC summaries,
mech-interp lesions) ship in the GitHub repo; this dataset holds the
policy checkpoints the CSVs were computed from.
Layout
runs/<group>__<reward>__A<xx>__seed_<n>.tar.zst— one archive per training run: final policies (policy_{seeker,hider}_final.pt), intermediate checkpoint history (checkpoints/{seeker,hider}_<upd>.pt, every ~50 updates, both roles),train.log, run metadata.finals/<group>/<reward>/A<xx>/seed_<n>/policy_*_final.pt— the final checkpoints loose, for single-policy downloads.
Groups: grid (the 2×2 factorial, n=20/cell), a_sweep (zoo dose
A ∈ {0.05, 0.1, 0.25}), coverage_ablation / urgency_ablation /
urgency_only_ablation (R7 term ablations), mixed_reward
(R7-seeker/R4-hider deconfound cells), gae_check (corrected-GAE
reruns), hsm_flank (hider-speed sensitivity), anchors (the 4
pre-registered evaluation anchors), pilot (three 10M-step pilots,
including the SAC HP-BROKEN pair).
Loading a policy
Policies are 87→128→128→(2·act) tanh MLPs saved as
{"pi": state_dict, "vf": state_dict}. With the GitHub repo checked
out:
from experiments.design_c_trajectory_analysis import load_policy
agent = load_policy("policy_seeker_final.pt", obs_dim=87, act_dim=2)
Provenance
Trained 2026-05/06 on LUMI and DTU gbar (LSF); rescued to this archive after LUMI's retirement. The study was pre-registered (v1–v3, frozen before data collection; see the results doc for the deviation ledger), survived an adversarial code-and-methods review, and the checkpoint histories here support the mechanistic analyses (input-block lesions, CKA, discovery-dynamics) described in Part IV of the results doc. Licensing follows the GitHub repository.
- Downloads last month
- 164