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
mask: string
opt: string
lr: double
lr_used: double
mask_size: int64
adam_first_step_norm: double
share_of_lr_sqrt_m: double
g1_norm: double
share_g_zero_%: double
share_g_below_1e-7_%: double
losses: list<item: double>
child 0, item: double
grad_norms: list<item: double>
child 0, item: double
step_norms: list<item: double>
child 0, item: double
steps: int64
loss_first50: double
loss_last50: double
R_LL_A_%: list<item: double>
child 0, item: double
R_LL_B_%: list<item: double>
child 0, item: double
R_LL_B_epoch1_%: list<item: double>
child 0, item: double
R_LL_B_tau0.9_%: list<item: double>
child 0, item: double
R_LL_B_tau0.8_%: list<item: double>
child 0, item: double
R_LL_B_tau0.7_%: list<item: double>
child 0, item: double
entropy_per_token_B: double
delta_norm: double
final_delta_norm_on_mask: double
cos_delta_vs_final: double
eff_delta_norm: double
cos_eff_delta_vs_final: double
visible_%: double
seq_lp5: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
seq_lp5_B_epoch1: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
secs_train: int64
secs: int64
final: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
base: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
to
{'base': List(List(Value('float64'))), 'final': List(List(Value('float64')))}
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
mask: string
opt: string
lr: double
lr_used: double
mask_size: int64
adam_first_step_norm: double
share_of_lr_sqrt_m: double
g1_norm: double
share_g_zero_%: double
share_g_below_1e-7_%: double
losses: list<item: double>
child 0, item: double
grad_norms: list<item: double>
child 0, item: double
step_norms: list<item: double>
child 0, item: double
steps: int64
loss_first50: double
loss_last50: double
R_LL_A_%: list<item: double>
child 0, item: double
R_LL_B_%: list<item: double>
child 0, item: double
R_LL_B_epoch1_%: list<item: double>
child 0, item: double
R_LL_B_tau0.9_%: list<item: double>
child 0, item: double
R_LL_B_tau0.8_%: list<item: double>
child 0, item: double
R_LL_B_tau0.7_%: list<item: double>
child 0, item: double
entropy_per_token_B: double
delta_norm: double
final_delta_norm_on_mask: double
cos_delta_vs_final: double
eff_delta_norm: double
cos_eff_delta_vs_final: double
visible_%: double
seq_lp5: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
seq_lp5_B_epoch1: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
secs_train: int64
secs: int64
final: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
base: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
to
{'base': List(List(Value('float64'))), 'final': List(List(Value('float64')))}
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.
Artifacts for "Fisher sparsity in RL post-training: does it depend on the optimizer?"
Code, report and analysis: https://github.com/jharkfinn/fisher-sparsity-rl
This repo holds the large artifacts of the OLMo 3 RL-Zero Math study, which tests arXiv 2609.36262 on a public RL run with intermediate checkpoints (allenai/Olmo-3-1025-7B → allenai/Olmo-3-7B-RL-Zero-Math).
| Path | Contents |
|---|---|
olmo3_rlzero/F_theta0_bf16/F_part*.pt |
Diagonal Fisher of allenai/Olmo-3-1025-7B with the protocol of arXiv 2609.36262 App. B.1 (eq. 17): 800 RL-Zero Math training prompts sampled without replacement (pandas seed 800), one on-policy response each (T = 1, top-p 1, no top-k, up to 4,096 tokens, vLLM), squared sequence-level score, FP32 parameters and accumulation, BF16 compute. Each tensor is the sum over the N = 800 responses (divide by N for eq. 17). Stored in BF16, keyed by HF parameter name, split into parts. meta.json gives N and the model. This copy is a recomputation from the same 800 responses after a compute-sandbox loss; it matches the first computation within 0.54 points on every concentration share and 0.014% in total mass. |
olmo3_rlzero/F_top1.pt |
The global top-1% set of that Fisher: per tensor, flat int32 indices and BF16 values, plus the quantile thresholds (thr) for the top 0.001–5%. |
olmo3_rlzero/resp.json |
The 800 prompts, prompt token ids and response token ids used for the Fisher (vLLM). |
olmo3_rlzero/resp_ckpt.json |
144 HF generate responses for the same prompts (sampler cross-check). |
olmo3_rlzero/results/*.json |
Concentration and structure of the Fisher (fcs_*; _rerun = the recomputation), the checkpoint delta analysis (olmo_delta.json), the Fisher-bin (Adam-ε) analysis (olmo_fbin.json) and the patch tests (olmo_patch_vllm_*). |
Derived from allenai/Olmo-3-1025-7B (Apache 2.0) and prompts from allenai/Dolci-RL-Zero-Math-7B; see those repos for their terms.
- Downloads last month
- 44,066