Datasets:

nick-rui
/

h-lens-artifacts

Tags:

Dataset Preview

Split (1)

test

The full dataset viewer is not available (click to read why). Only showing a preview of the rows.

The dataset generation failed because of a cast error

Error code:   DatasetGenerationCastError
Exception:    DatasetGenerationCastError
Message:      An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 3 new columns ({'k', 'weighting', 'pass_at_k'}) and 5 missing columns ({'rank', 'item', 'layer', 'intermediate', 'n_candidates'}).

This happened while the csv dataset builder was generating data using

hf://datasets/nick-rui/h-lens-artifacts/hc/eval_pool_all_pass_at_k.csv (at revision 6173b20f6df6ac7a9d8812dca3425630adbcbc32), ['hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_pool_all.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_pool_all_pass_at_k.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit128.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit128_pass_at_k.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit_labels.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit_labels_pass_at_k.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit_pass_at_k.csv']

Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1848, in _prepare_split_single
                  writer.write_table(table)
                  ~~~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 784, in write_table
                  self._write_table(pa_table, writer_batch_size=writer_batch_size)
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 795, in _write_table
                  pa_table = table_cast(pa_table, self._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
              eval: string
              lens: string
              k: int64
              weighting: string
              pass_at_k: double
              -- schema metadata --
              pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 830
              to
              {'eval': Value('string'), 'lens': Value('string'), 'item': Value('string'), 'intermediate': Value('string'), 'layer': Value('int64'), 'rank': Value('int64'), 'n_candidates': Value('int64')}
              because column names don't match
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1850, in _prepare_split_single
                  raise DatasetGenerationCastError.from_cast_error(
                  ...<4 lines>...
                  )
              datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
              
              All the data files must have the same columns, but at some point there are 3 new columns ({'k', 'weighting', 'pass_at_k'}) and 5 missing columns ({'rank', 'item', 'layer', 'intermediate', 'n_candidates'}).
              
              This happened while the csv dataset builder was generating data using
              
              hf://datasets/nick-rui/h-lens-artifacts/hc/eval_pool_all_pass_at_k.csv (at revision 6173b20f6df6ac7a9d8812dca3425630adbcbc32), ['hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_pool_all.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_pool_all_pass_at_k.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit128.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit128_pass_at_k.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit_labels.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit_labels_pass_at_k.csv', 'hf://datasets/nick-rui/h-lens-artifacts@6173b20f6df6ac7a9d8812dca3425630adbcbc32/hc/eval_refit_pass_at_k.csv']
              
              Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)

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.

eval string	lens string	item string	intermediate string	layer int64	rank int64	n_candidates int64
multihop	lens_pool_all	carnival-ocean	Brazil	8	18,085	3
multihop	lens_pool_all	carnival-ocean	Brazil	16	6,655	3
multihop	lens_pool_all	carnival-ocean	Brazil	24	1,987	3
multihop	lens_pool_all	carnival-ocean	Brazil	32	1,086	3
multihop	lens_pool_all	carnival-ocean	Brazil	40	400	3
multihop	lens_pool_all	carnival-ocean	Brazil	48	305	3
multihop	lens_pool_all	carnival-ocean	Brazil	56	18	3
multihop	jpp	carnival-ocean	Brazil	8	180,185	3
multihop	jpp	carnival-ocean	Brazil	16	26,056	3
multihop	jpp	carnival-ocean	Brazil	24	3,271	3
multihop	jpp	carnival-ocean	Brazil	32	1,296	3
multihop	jpp	carnival-ocean	Brazil	40	425	3
multihop	jpp	carnival-ocean	Brazil	48	439	3
multihop	jpp	carnival-ocean	Brazil	56	40	3
multihop	lens_pool_all	amazon-language	Brazil	8	1,931	3
multihop	lens_pool_all	amazon-language	Brazil	16	1,132	3
multihop	lens_pool_all	amazon-language	Brazil	24	199	3
multihop	lens_pool_all	amazon-language	Brazil	32	420	3
multihop	lens_pool_all	amazon-language	Brazil	40	42	3
multihop	lens_pool_all	amazon-language	Brazil	48	41	3
multihop	lens_pool_all	amazon-language	Brazil	56	4	3
multihop	jpp	amazon-language	Brazil	8	7,886	3
multihop	jpp	amazon-language	Brazil	16	3,346	3
multihop	jpp	amazon-language	Brazil	24	188	3
multihop	jpp	amazon-language	Brazil	32	421	3
multihop	jpp	amazon-language	Brazil	40	41	3
multihop	jpp	amazon-language	Brazil	48	40	3
multihop	jpp	amazon-language	Brazil	56	4	3
multihop	lens_pool_all	mars-color	Mars	8	819	3
multihop	lens_pool_all	mars-color	Mars	16	1,846	3
multihop	lens_pool_all	mars-color	Mars	24	59	3
multihop	lens_pool_all	mars-color	Mars	32	146	3
multihop	lens_pool_all	mars-color	Mars	40	7,030	3
multihop	lens_pool_all	mars-color	Mars	48	1,127	3
multihop	lens_pool_all	mars-color	Mars	56	18	3
multihop	jpp	mars-color	Mars	8	610	3
multihop	jpp	mars-color	Mars	16	770	3
multihop	jpp	mars-color	Mars	24	54	3
multihop	jpp	mars-color	Mars	32	241	3
multihop	jpp	mars-color	Mars	40	4,635	3
multihop	jpp	mars-color	Mars	48	841	3
multihop	jpp	mars-color	Mars	56	14	3
multihop	lens_pool_all	spider-legs	spider	8	14,435	3
multihop	lens_pool_all	spider-legs	spider	16	5,807	3
multihop	lens_pool_all	spider-legs	spider	24	5	3
multihop	lens_pool_all	spider-legs	spider	32	6	3
multihop	lens_pool_all	spider-legs	spider	40	1	3
multihop	lens_pool_all	spider-legs	spider	48	1	3
multihop	lens_pool_all	spider-legs	spider	56	1	3
multihop	jpp	spider-legs	spider	8	159,333	3
multihop	jpp	spider-legs	spider	16	10,350	3
multihop	jpp	spider-legs	spider	24	5	3
multihop	jpp	spider-legs	spider	32	6	3
multihop	jpp	spider-legs	spider	40	1	3
multihop	jpp	spider-legs	spider	48	1	3
multihop	jpp	spider-legs	spider	56	1	3
multihop	lens_pool_all	paper-continent	China	8	4,031	4
multihop	lens_pool_all	paper-continent	China	16	659	4
multihop	lens_pool_all	paper-continent	China	24	117	4
multihop	lens_pool_all	paper-continent	China	32	161	4
multihop	lens_pool_all	paper-continent	China	40	1	4
multihop	lens_pool_all	paper-continent	China	48	3	4
multihop	lens_pool_all	paper-continent	China	56	4	4
multihop	jpp	paper-continent	China	8	12,944	4
multihop	jpp	paper-continent	China	16	951	4
multihop	jpp	paper-continent	China	24	154	4
multihop	jpp	paper-continent	China	32	133	4
multihop	jpp	paper-continent	China	40	2	4
multihop	jpp	paper-continent	China	48	3	4
multihop	jpp	paper-continent	China	56	5	4
multihop	lens_pool_all	atomic-79-symbol	gold	8	5	4
multihop	lens_pool_all	atomic-79-symbol	gold	16	35	4
multihop	lens_pool_all	atomic-79-symbol	gold	24	104	4
multihop	lens_pool_all	atomic-79-symbol	gold	32	723	4
multihop	lens_pool_all	atomic-79-symbol	gold	40	12	4
multihop	lens_pool_all	atomic-79-symbol	gold	48	98	4
multihop	lens_pool_all	atomic-79-symbol	gold	56	12	4
multihop	jpp	atomic-79-symbol	gold	8	7	4
multihop	jpp	atomic-79-symbol	gold	16	8	4
multihop	jpp	atomic-79-symbol	gold	24	103	4
multihop	jpp	atomic-79-symbol	gold	32	558	4
multihop	jpp	atomic-79-symbol	gold	40	10	4
multihop	jpp	atomic-79-symbol	gold	48	74	4
multihop	jpp	atomic-79-symbol	gold	56	5	4
multihop	lens_pool_all	atomic-29-symbol	copper	8	14	2
multihop	lens_pool_all	atomic-29-symbol	copper	16	89	2
multihop	lens_pool_all	atomic-29-symbol	copper	24	22	2
multihop	lens_pool_all	atomic-29-symbol	copper	32	66	2
multihop	lens_pool_all	atomic-29-symbol	copper	40	7	2
multihop	lens_pool_all	atomic-29-symbol	copper	48	115	2
multihop	lens_pool_all	atomic-29-symbol	copper	56	10	2
multihop	jpp	atomic-29-symbol	copper	8	23	2
multihop	jpp	atomic-29-symbol	copper	16	1,433	2
multihop	jpp	atomic-29-symbol	copper	24	25	2
multihop	jpp	atomic-29-symbol	copper	32	60	2
multihop	jpp	atomic-29-symbol	copper	40	5	2
multihop	jpp	atomic-29-symbol	copper	48	129	2
multihop	jpp	atomic-29-symbol	copper	56	6	2
multihop	lens_pool_all	month-3-godof	March	8	80	3
multihop	lens_pool_all	month-3-godof	March	16	298	3

End of preview.

H-lens artifacts: Qwen3.6-27B

These are intermediate outputs of the H-lens mini-project. The project adds a context-averaged second-order term (a "Hessian wrapper") to the workspace lenses J-lens, R-lens and J++. All outputs here are for Qwen/Qwen3.6-27B, revision 6a9e13bd. The Qwen3.5-9B outputs are not included.

The code and the write-up live in the h-lens repository, mainly its results/RESULTS.md. The file names below are the ones that code reads from its OUT directory, so a clone can point OUT at a download of this repository.

Files	Contents
`t2_fwd_scale_<tag>[_s<i>].pt`	Per-context first and second directional derivatives. `J1` and `H2` have shape [items, layers, K, d]: for each eval item and read layer, over pile contexts (target block n−2). Written by `code/t2_fwd_scale2.py`.
`t2_fwd_hpp_P27*.pt`	The same on J++'s fitting positions, stratified by its router (H++). Written by `code/t2_fwd_hpp.py`.
`t2_split_S27*.pt`	The near part of the near/far split of the second-order term (`Hnear`). Written by `code/t2_fwd_split.py`.
`trad27_resid.pt`, `wsb*_resid.pt`, `hc/resid.pt`, `hc/wsb_resid.pt`	Residual captures of the eval items: the J++ held-out:0 tasks and the single-token WorkspaceBench families.
`hc/lens_refit.pt`, `hc/router_K.pt`, `hc/*expert_weights.json`	Our J++ refit (jpp_lens recipe, 64 or 128 WikiText prompts), with its routers and expert weights.
`hc/experts.pt`, `hc/pool_.pt`, `hc/bag/`	Per-expert and pooled Jacobian maps, and bagged refits.
`hc/artifacts/`	The refit's shard artifacts, needed only to refit new variants.
`hc/hmeans_*.pt`	Per-read context means of the derivative files.
`e0/jpp.pt`	The released J++ map (koayon/jpp-lenses), converted.
`unembed.pt`	The final norm and the unembedding.

The tag says which graph the derivatives were taken through:

A27: plain gradients (H-lens).
R27: the R-lens LRP graph (RH-lens).
Q27 / F1_27: the R-lens graph plus a freeze of the q/k norms.
F2_27: also freezes the attention pattern.
F3_27: F2 plus the output gates.
FM_27: the MLP product linearized. Partial: held-out only, and FM_27bas has layers 56–16 (see layers_done).

The suffix says which eval part the file covers:

none: J++ held-out:0;
w, bas, brw, dm: WorkspaceBench parts;
b, c: more contexts for the same items (k0).

External code: safety-research/jpp_lens @ 93f02cd and camilablank/workspace-bench @ 92d763e.

Downloads last month: 2,308

Total file size:

613 GB