Dataset Viewer
Duplicate
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
artifact: string
source_repo: string
source_commit: string
contents: list<item: string>
  child 0, item: string
training_performed: bool
models_or_credentials: bool
intended_consumer: string
phase: string
student: struct<repo: string, mode: string>
  child 0, repo: string
  child 1, mode: string
provenance: struct<runtime: string, author_equivalence: string, training_performed: bool>
  child 0, runtime: string
  child 1, author_equivalence: string
  child 2, training_performed: bool
hardware: struct<minimum_b200: int64, parallel_b200: int64>
  child 0, minimum_b200: int64
  child 1, parallel_b200: int64
status: string
schema_version: int64
training_data: struct<repo: string, rows: int64>
  child 0, repo: string
  child 1, rows: int64
phase0: struct<rollout_cap: int64, planned_semi_steps: int64, planned_opd_steps: int64, checkpoints: list<it (... 139 chars omitted)
  child 0, rollout_cap: int64
  child 1, planned_semi_steps: int64
  child 2, planned_opd_steps: int64
  child 3, checkpoints: list<item: int64>
      child 0, item: int64
  child 4, pool_prompts: int64
  child 5, held_out_rollouts: int64
  child 6, fresh_rollouts_every_steps: int64
  child 7, fresh_rollout_count: int64
  child 8, dev_eval: string
objectives: list<item: string>
  child 0, item: string
teacher: struct<repo: string>
  child 0, repo: string
project: string
to
{'schema_version': Value('int64'), 'project': Value('string'), 'phase': Value('string'), 'status': Value('string'), 'hardware': {'minimum_b200': Value('int64'), 'parallel_b200': Value('int64')}, 'student': {'repo': Value('string'), 'mode': Value('string')}, 'teacher': {'repo': Value('string')}, 'training_data': {'repo': Value('string'), 'rows': Value('int64')}, 'objectives': List(Value('string')), 'phase0': {'rollout_cap': Value('int64'), 'planned_semi_steps': Value('int64'), 'planned_opd_steps': Value('int64'), 'checkpoints': List(Value('int64')), 'pool_prompts': Value('int64'), 'held_out_rollouts': Value('int64'), 'fresh_rollouts_every_steps': Value('int64'), 'fresh_rollout_count': Value('int64'), 'dev_eval': Value('string')}, 'provenance': {'runtime': Value('string'), 'author_equivalence': Value('string'), 'training_performed': Value('bool')}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, 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 129, 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 544, 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 3052, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, 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 558, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, 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 361, 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 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              artifact: string
              source_repo: string
              source_commit: string
              contents: list<item: string>
                child 0, item: string
              training_performed: bool
              models_or_credentials: bool
              intended_consumer: string
              phase: string
              student: struct<repo: string, mode: string>
                child 0, repo: string
                child 1, mode: string
              provenance: struct<runtime: string, author_equivalence: string, training_performed: bool>
                child 0, runtime: string
                child 1, author_equivalence: string
                child 2, training_performed: bool
              hardware: struct<minimum_b200: int64, parallel_b200: int64>
                child 0, minimum_b200: int64
                child 1, parallel_b200: int64
              status: string
              schema_version: int64
              training_data: struct<repo: string, rows: int64>
                child 0, repo: string
                child 1, rows: int64
              phase0: struct<rollout_cap: int64, planned_semi_steps: int64, planned_opd_steps: int64, checkpoints: list<it (... 139 chars omitted)
                child 0, rollout_cap: int64
                child 1, planned_semi_steps: int64
                child 2, planned_opd_steps: int64
                child 3, checkpoints: list<item: int64>
                    child 0, item: int64
                child 4, pool_prompts: int64
                child 5, held_out_rollouts: int64
                child 6, fresh_rollouts_every_steps: int64
                child 7, fresh_rollout_count: int64
                child 8, dev_eval: string
              objectives: list<item: string>
                child 0, item: string
              teacher: struct<repo: string>
                child 0, repo: string
              project: string
              to
              {'schema_version': Value('int64'), 'project': Value('string'), 'phase': Value('string'), 'status': Value('string'), 'hardware': {'minimum_b200': Value('int64'), 'parallel_b200': Value('int64')}, 'student': {'repo': Value('string'), 'mode': Value('string')}, 'teacher': {'repo': Value('string')}, 'training_data': {'repo': Value('string'), 'rows': Value('int64')}, 'objectives': List(Value('string')), 'phase0': {'rollout_cap': Value('int64'), 'planned_semi_steps': Value('int64'), 'planned_opd_steps': Value('int64'), 'checkpoints': List(Value('int64')), 'pool_prompts': Value('int64'), 'held_out_rollouts': Value('int64'), 'fresh_rollouts_every_steps': Value('int64'), 'fresh_rollout_count': Value('int64'), 'dev_eval': Value('string')}, 'provenance': {'runtime': Value('string'), 'author_equivalence': Value('string'), 'training_performed': Value('bool')}}
              because column names don't match

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.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Speed Limit of Distillation

Phase 0 implementation of the experiment plan, using Qwen3-1.7B non-thinking, Qwen3-4B-Instruct-2507 and the pinned 14,116-row DAPO input. This is an independent Transformers/vLLM reconstruction using the existing NeMo-RL v0.7 runtime dependencies. Equivalence to the authors' training backend has not been established. A running job is not a reproduction pass.

The three single-GPU queues run:

  1. Numerical tests and a two-step real-model canary, then 10 OPD calibration steps with separate generation, teacher-scoring and optimizer timings.
  2. Two disjoint cache shards, covering every prompt once with a genuine initial student rollout (4,096-token cap), sampled teacher logp and FP32 teacher top-64 logp plus token IDs at every response position.
  3. Semi-OPD, 100 steps, global batch 128, constant LR 1e-6, AdamW (0.9, 0.999), weight decay .01 and gradient clipping 1. FP32 master weights and optimizer state, BF16 computation; one sequence per gradient accumulation.
  4. Independently queued full final evaluations at checkpoints 60/80/100, AIME24/25/26 + HMMT26, avg@32, 32,768-token cap. Training need not wait for final evaluation. The aggregate pass gate uses the equal-benchmark average and then the equal-checkpoint average, compared with .398 +/- .03.

The cache creates a fixed 2,048-prompt pool, a disjoint 256-prompt held-out set and three seeded 64-rollout subsets. These held-out prompts are disjoint from pool P for later repeated-S experiments; Phase 0's global Semi reproduction still trains on all training prompts. Diagnostic KL uses teacher top-64 candidate sets, independently renormalized teacher/student distributions, weighted equally over token states. Log sampled-token training loss separately. The diagnostic global-training curve must not be called a repeated-S curve. Measurements include step zero, every 10 updates, held-out KL and fresh 128 held-out rollouts at steps 50/100. Missing threshold crossings are censored; never replace a missing crossing with a measured time.

speed_limit.lanczos implements a matrix-free VJP/JVP loss-gradient Gram operator, Adam second-moment preconditioning and reorthogonalized Lanczos. The numerical test compares with explicit tiny-model gradients and eigenvalues at 1% tolerance. A LoRA/full-training equivalence check and a validated spectral prediction are future research gates; a passing toy test does not establish them.

Running with the existing remote runtime

No local model downloads or environment builds are needed. On each selected node, use its existing runtime interpreter and set PYTHONPATH to this repo's src. Each role receives one physical GPU UUID; the queue refuses a GPU with compute processes or more than 1 GiB allocated. The controller resolves that UUID to a numeric index for vLLM and each model process verifies the CUDA-visible UUID against the requested physical GPU before loading any model. Provide fresh run/source paths.

PYTHONPATH="$SOURCE/src" "$INPUTS/runtime/envs/vllm/bin/python" \
  -m speed_limit.queue --source "$SOURCE" --inputs "$INPUTS" \
  --run "$RUN" --role 0 --gpu "$GPU_UUID" --revision "$GIT_COMMIT" --detach

Start roles 0, 1, 2, exactly once each, sharing the same run directory. Role 0 owns calibration and training; roles 1/2 own cache shards 0/1 and final evals (60/100 and 80). Controllers detach from Jupyter, preserve task logs, record child PIDs and heartbeat state, and stop their queue on failures. Unknown Bridge commands are never retried. No Phase 1 work launches automatically.

Offline numerical checks (inside the existing runtime):

PYTHONPATH=src python -m unittest discover -s tests -v

The source bundle contains no models, datasets, runtime, credentials or run results. Publish the verified Git commit first, archive it, upload the archive and its SHA-256 manifest to Hugging Face, then fetch the pinned HF revision on nodes through the supplied proxy. Source and runtime have separate provenance.

A read-only evidence aggregator is available as src/speed_limit/report.py RUN. It reports partial progress explicitly and separates cache generation, teacher scoring, refit and updates. --watch --output PATH --detach maintains an atomic summary on shared storage until the queue completes or a failure marker appears. A newer reporting overlay can inspect a run frozen to an older training commit; both source revisions remain in their respective receipts.

When the proxy cannot reach the HF binary CDN, source-bundle.json transports the same Git archive as base64 through HF's regular text-file endpoint. The deployer validates both Git revision and decoded archive SHA-256 before use.

Validate actual saved cache shards without loading a GPU model:

PYTHONPATH=src python -m speed_limit.cache --file /path/to/batch-00000.pt --expected-rows 32

This checks weights-only tensor deserialization, token alignment, candidate probability mass and agreement between sampled logp and matching top-64 IDs.

Move a checkpoint evaluation to another allocated GPU with the detached launcher:

python scripts/run_checkpoint.py --inputs "$INPUTS" --runner-source "$RUNNER_SOURCE" \
  --training-run "$RUN" --output "$ATTEMPT" --step 100 --gpu "$GPU_UUID" \
  --revision "$LAUNCHER_COMMIT" --detach

The launcher reuses the qualified runner's environment, checks the checkpoint ready marker and GPU availability, and records process, progress and completion receipts in the fresh attempt directory. Add --dev for AIME24/25 avg@8. When moving a job previously scheduled by watch-and-launch-step100.sh, use --retire-watch-run RUN --retirement-receipt FRESH_JSON on that watcher's host. The retirement checks the PID's command line before stopping the owned watcher. To continue a stopped evaluator, add --prior-result PRIOR_RESULT_DIR. The launcher validates the previous run settings and reuses complete four-prompt batches. A partially written batch is generated again with the original seed.

Concurrent exploratory K1 training

speed_limit.fixed repeats updates only on one of the three cached 64-rollout sets from splits.json. The initial student, AdamW settings and sampled-token sequence-sum loss match Setup. It measures token-weighted top64 KL on S at step zero and every update, held-out KL every 10 steps, and 128 fresh held-out rollouts every 50 steps and at termination. Fresh rollouts are diagnostics. The run ends at 300 updates or the first measured A below .1.

scripts/stack_k1.py runs one set on each of roles 1/2/3, sweeping LR 1e-6, 3e-6, 1e-5, 3e-5 and 1e-4. Each run starts from the initial student. Role 0 waits for the relocated Phase0 step100 evaluation, evaluates its dev set, then serves the dev evaluation requests from completed K1 runs serially. The user authorized this overlap while the Phase0 final benchmark gate is pending. Run receipts retain that pending status. K1 interpretation follows verification of the Phase0 model, objective and evaluation protocol.

Fixed-set checkpoints at steps 50, 100, 200 and 300, or early termination, include model, optimizer and Python/NumPy/Torch/CUDA RNG state. Resume into a fresh output directory with python -m speed_limit.fixed ... --resume CHECKPOINT. The runner compares the saved set IDs, LR, optimizer and cache manifests, restores optimizer/RNG state, and links the earlier absorption history. The optimizer option also supports SGD with momentum .9 for the subsequent best-two-LR comparison after the AdamW sweep is analyzed.

K1 dev outputs and logs live under the campaign dev/ directory. The GPU-0 queue state records completion paths for every attempted dev evaluation. To relocate a sweep, start a fresh campaign with --previous-run PRIOR_CAMPAIGN. The training controller reuses completed runs and resumes incomplete runs from their latest ready checkpoint. Pass --training-roles to the evaluation controller when the destination campaign contains only a subset of the original roles.

Bounded multi-node campaign

scripts/campaign.py prepares 15 K1 AdamW runs and five exploratory S0 top64 LR runs. Each GPU worker owns an isolated output tree and claims its assigned jobs exclusively. Recovery workers can select a prior configuration by index, then run deferred pilot configurations. --pilot-set extends a top64 pilot to another fixed state set while preserving its LR and objective. Each completed training run queues AIME24/25 dev evaluation. After all 15 K1 dev results, mean dev accuracy across S0/S1/S2 selects two LRs for six SGD runs. Ties prefer the smaller LR. The top dev LR also schedules top64 on S1/S2. This ranking is provisional until held-out KL and fresh-rollout length are reviewed. The vOPD control-variate arm and real-model K3 spectra remain pending. Workers accept one physical GPU UUID and an absolute UTC deadline. Each payload runs in its own session under GNU timeout with TERM and a five-second KILL grace, so payloads remain bounded if the controller disappears. Controllers stop launching ten seconds before the deadline. Resume uses prior ready checkpoints with the saved model, optimizer and RNG state. Source and output paths use the relocated /workspace/storage-shared/nlp/users/tungks root.

Downloads last month
83