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
artifact_sha256: struct<4: string>
child 0, 4: string
bits: list<item: int64>
child 0, item: int64
calibration_freeze_sha256: string
calibration_report: struct<4: struct<bos_trace_share: struct<max: double, max_target: string, mean: double>, dead_total: (... 61 chars omitted)
child 0, 4: struct<bos_trace_share: struct<max: double, max_target: string, mean: double>, dead_total: int64, n_ (... 50 chars omitted)
child 0, bos_trace_share: struct<max: double, max_target: string, mean: double>
child 0, max: double
child 1, max_target: string
child 2, mean: double
child 1, dead_total: int64
child 2, n_rows: int64
child 3, peak_gpu_gb: double
child 4, seconds: double
code_sha256: struct<__init__.py: string, artifact.py: string, calibrate.py: string, calibration.py: string, canon (... 183 chars omitted)
child 0, __init__.py: string
child 1, artifact.py: string
child 2, calibrate.py: string
child 3, calibration.py: string
child 4, canonical.py: string
child 5, compare_stage1.py: string
child 6, gate_run.py: string
child 7, grids.py: string
child 8, run_with_weights.py: string
child 9, states.py: string
child 10, upstream/gptq.py: string
child 11, upstream/quant.py: string
config: struct<actorder: bool, bits: string, blocksize: int64, groupsize: int64, mse: bool, percdamp: double (... 12 chars omitted)
child 0, actorder: bool
child 1, bits: string
child 2, blocksize: int64
child 3, groupsize: int64
child 4, mse: bool
child 5, percdamp: double
child 6, sym: bool
gptq_freeze_sha256: string
method: string
pool_sha256: string
preset: string
model: string
skipped_for_time: list<item: null>
child 0, item: null
gates: struct<split_equals_full: bool, scale_one_identity: bool, donor_self_identity: bool, sham_identity: (... 5 chars omitted)
child 0, split_equals_full: bool
child 1, scale_one_identity: bool
child 2, donor_self_identity: bool
child 3, sham_identity: bool
argv: list<item: string>
child 0, item: string
commit: string
peak_gpu_gb: double
task: string
n_sentences: int64
seconds: double
split: string
languages: list<item: string>
child 0, item: string
precisions: list<item: string>
child 0, item: string
manifest_source_sha256: string
complete: bool
hook: string
versions: struct<torch: string, transformer-lens: string, sae-lens: string, transformers: string, datasets: st (... 5 chars omitted)
child 0, torch: string
child 1, transformer-lens: string
child 2, sae-lens: string
child 3, transformers: string
child 4, datasets: string
summary: struct<INT4: struct<seconds: double, token_passes: int64, tokens_per_second: double, configs: int64> (... 1 chars omitted)
child 0, INT4: struct<seconds: double, token_passes: int64, tokens_per_second: double, configs: int64>
child 0, seconds: double
child 1, token_passes: int64
child 2, tokens_per_second: double
child 3, configs: int64
lm_seed: int64
to
{'task': Value('string'), 'split': Value('string'), 'preset': Value('string'), 'precisions': List(Value('string')), 'languages': List(Value('string')), 'n_sentences': Value('int64'), 'lm_seed': Value('int64'), 'commit': Value('string'), 'model': Value('string'), 'hook': Value('string'), 'gates': {'split_equals_full': Value('bool'), 'scale_one_identity': Value('bool'), 'donor_self_identity': Value('bool'), 'sham_identity': Value('bool')}, 'manifest_source_sha256': Value('string'), 'argv': List(Value('string')), 'versions': {'torch': Value('string'), 'transformer-lens': Value('string'), 'sae-lens': Value('string'), 'transformers': Value('string'), 'datasets': Value('string')}, 'summary': {'INT4': {'seconds': Value('float64'), 'token_passes': Value('int64'), 'tokens_per_second': Value('float64'), 'configs': Value('int64')}}, 'skipped_for_time': List(Value('null')), 'complete': Value('bool'), 'seconds': Value('float64'), 'peak_gpu_gb': 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
artifact_sha256: struct<4: string>
child 0, 4: string
bits: list<item: int64>
child 0, item: int64
calibration_freeze_sha256: string
calibration_report: struct<4: struct<bos_trace_share: struct<max: double, max_target: string, mean: double>, dead_total: (... 61 chars omitted)
child 0, 4: struct<bos_trace_share: struct<max: double, max_target: string, mean: double>, dead_total: int64, n_ (... 50 chars omitted)
child 0, bos_trace_share: struct<max: double, max_target: string, mean: double>
child 0, max: double
child 1, max_target: string
child 2, mean: double
child 1, dead_total: int64
child 2, n_rows: int64
child 3, peak_gpu_gb: double
child 4, seconds: double
code_sha256: struct<__init__.py: string, artifact.py: string, calibrate.py: string, calibration.py: string, canon (... 183 chars omitted)
child 0, __init__.py: string
child 1, artifact.py: string
child 2, calibrate.py: string
child 3, calibration.py: string
child 4, canonical.py: string
child 5, compare_stage1.py: string
child 6, gate_run.py: string
child 7, grids.py: string
child 8, run_with_weights.py: string
child 9, states.py: string
child 10, upstream/gptq.py: string
child 11, upstream/quant.py: string
config: struct<actorder: bool, bits: string, blocksize: int64, groupsize: int64, mse: bool, percdamp: double (... 12 chars omitted)
child 0, actorder: bool
child 1, bits: string
child 2, blocksize: int64
child 3, groupsize: int64
child 4, mse: bool
child 5, percdamp: double
child 6, sym: bool
gptq_freeze_sha256: string
method: string
pool_sha256: string
preset: string
model: string
skipped_for_time: list<item: null>
child 0, item: null
gates: struct<split_equals_full: bool, scale_one_identity: bool, donor_self_identity: bool, sham_identity: (... 5 chars omitted)
child 0, split_equals_full: bool
child 1, scale_one_identity: bool
child 2, donor_self_identity: bool
child 3, sham_identity: bool
argv: list<item: string>
child 0, item: string
commit: string
peak_gpu_gb: double
task: string
n_sentences: int64
seconds: double
split: string
languages: list<item: string>
child 0, item: string
precisions: list<item: string>
child 0, item: string
manifest_source_sha256: string
complete: bool
hook: string
versions: struct<torch: string, transformer-lens: string, sae-lens: string, transformers: string, datasets: st (... 5 chars omitted)
child 0, torch: string
child 1, transformer-lens: string
child 2, sae-lens: string
child 3, transformers: string
child 4, datasets: string
summary: struct<INT4: struct<seconds: double, token_passes: int64, tokens_per_second: double, configs: int64> (... 1 chars omitted)
child 0, INT4: struct<seconds: double, token_passes: int64, tokens_per_second: double, configs: int64>
child 0, seconds: double
child 1, token_passes: int64
child 2, tokens_per_second: double
child 3, configs: int64
lm_seed: int64
to
{'task': Value('string'), 'split': Value('string'), 'preset': Value('string'), 'precisions': List(Value('string')), 'languages': List(Value('string')), 'n_sentences': Value('int64'), 'lm_seed': Value('int64'), 'commit': Value('string'), 'model': Value('string'), 'hook': Value('string'), 'gates': {'split_equals_full': Value('bool'), 'scale_one_identity': Value('bool'), 'donor_self_identity': Value('bool'), 'sham_identity': Value('bool')}, 'manifest_source_sha256': Value('string'), 'argv': List(Value('string')), 'versions': {'torch': Value('string'), 'transformer-lens': Value('string'), 'sae-lens': Value('string'), 'transformers': Value('string'), 'datasets': Value('string')}, 'summary': {'INT4': {'seconds': Value('float64'), 'token_passes': Value('int64'), 'tokens_per_second': Value('float64'), 'configs': Value('int64')}}, 'skipped_for_time': List(Value('null')), 'complete': Value('bool'), 'seconds': Value('float64'), 'peak_gpu_gb': 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.
always-late: large run artifacts
Large artifacts of the ANLP course project "Does compression forget language?" (team Always Late, IIIT-H). Code, documentation and small summaries live in the project repository TH4KUR/always-late (private at the time of writing; its reports/artifacts.md indexes this dataset and reports/audit_our_vs_duan.md explains the baseline runs). This dataset holds what is too large for git: per-feature tables, per-language statistics, bootstrap snapshots and token windows. Everything here is a result of a course project and is provisional; read the status notes below before reusing a number.
Contents are derived statistics only (per-feature correlations, firing rates and activation summaries of the Gemma Scope layer-12 16k SAE, perplexities, token ids of WikiText-2 windows). No model weights, no SAE weights, no generated text and no FLORES-200 sentences or token ids.
Licences and attribution
- Our own statistics: the team has not chosen a licence for them (
license: other). Cite the project repository and this dataset; the terms of the sources below still apply. - Gemma-2-2B and Gemma Scope (
google/gemma-2-2b,google/gemma-scope-2b-pt-res) were used under their own terms; none of their weights are redistributed here. - WikiText-2 (Merity et al., 2016; text from Wikipedia, CC BY-SA): the
tokens_2d.ptfiles hold token ids of WikiText-2 test text and are therefore an adaptation of it. Use those files under CC BY-SA with attribution to the WikiText-2 authors and Wikipedia contributors. - FLORES-200 (NLLB Team, CC BY-SA 4.0): the Phase 1 folders contain statistics computed from its
devanddevtestsplits, not its text. - Duan (2026) (arXiv 2606.03002, MIT-licensed code):
baseline/,smoke/,runs/R1-*andruns/R2-*were produced by running their vendored script on WikiText-2. Their paper is a preprint (submitted to TMLR when last checked).
Contents
| folder | what | Slurm job | code |
|---|---|---|---|
baseline/full |
Duan et al. (2026) Phase-3 script, vendored, 500k WikiText-2 tokens (256-token windows, no BOS), INT8..INT4 | 23472 | third_party/duan2026 (working.sha256 7f1e638a0508c3ca) |
smoke/R0-v3-runner, smoke/generator-runner, smoke/vendored |
20k-token smokes (INT8, INT6) | 23355, 23419, 23463 | diagnostic generator / removed faithful runner / vendored script |
runs/R1-nobos-floor |
their code, no BOS, INT0 / INT16 / INT12 (noise floor) | 24451 | scripts/g0/make_duan_runners_v3.py, runner hash 2c34fecaee1d2e32 |
runs/R2-bos-mask |
their code with BOS prepended and masked, INT0 / 16 / 12 / 8 / 7 / 6 | 23643 | same generator |
runs/R5a, runs/R5b, runs/R5c |
the discarded HF harness (git tag legacy-pipeline-2026-10-16) on their tokens, one factor flipped per cell: R5a = T, O, T_sdpa, T_bf16sae; R5b = T_rowquant, T_mask, T_keep, O_keep; R5c = O_mask |
23649, 23660, 24452 | scripts/g0/r5_run.py (tag), hashes in each code.sha256 |
runs/phase1/p1-mask, runs/phase1/p1-none |
Phase 1 multilingual run: 27 FLORES-200 languages (dev split for feature tagging, devtest split for survival), Gemma-2-2B layer-12 16k SAE, RTN INT8/7/6/5/4 plus INT16/12 and the restore-only control INT0. p1-mask prepends BOS and excludes its position from every statistic (the SAE's own convention; the primary regime). p1-none has no BOS and is a historical regime that failed its noise-floor check. Each holds eq/ (equal token budget per language) and all/ (all tokens) |
25745, 25746 | scripts/g0/phase1_run.py at commit 618496f |
runs/phase1/association-full |
D-035 association-retention run: the original mask/eq cohort (8,086 language-selective features and 2,860 matched controls, frozen from p1-mask) evaluated on FLORES-200 devtest (27 languages, 1,012 aligned sentences each), full precision and RTN INT8/7/6/5/4, INT16/12 and the restore-only control INT0, mask regime, with each condition's OWN quantized-side language profiles: association/{eq,all}/profiles.pt (per condition, language and feature: positive-token counts, activation sums and maxima, content tokens), the per-sentence shards association/eq/sentences/<condition>/<language>.pt (totals of the 10,946 selective-plus-control features as CSR rows, float32 sums, equal-budget stream, 243 files, about 2 GB) are NOT in this dataset (the uplink of the cluster was too slow; they stay on the cluster and are regenerated by the recorded run command, and their hashes are in association/manifest.json), association/manifest.json (hashes and sizes); eq/ and all/ hold the usual legacy outputs (identical to p1-mask); analysis-eq/ (1,000 paired sentence-bootstrap draws, seed 0), analysis-all/ (aggregate profile only) and crosscheck-eq/ hold the generated reports (analysis.json, feature_status.csv, language_summary.csv, association_retention.md, crosscheck.*) |
31640 (run), 31807 / 31808 (analyses) | scripts/g0/phase1_run.py --association-profiles ... (commit 1655450), scripts/g0/phase1_association_analysis.py and phase1_association_crosscheck.py (commits dc7c3e1, 3496916) |
runs/phase1/anchor-batch1, anchor-batch4 |
code-regression check of the Phase 1 runner against the vendored smoke run | see repository | scripts/g0/phase1_anchor.py |
runs/phase1/smoke-flores-3lang, smoke2-flores-3lang-eq |
3-language smokes of the Phase 1 runner | see repository | scripts/g0/phase1_run.py |
File types: *_per_feature.csv (feature_id, condition, layer, corr, firing_rate, mean_activation, max_activation, active, survived, damaged: Duan's schema), *_summary.csv / phase3_summary_final.csv, weight_perturbation_INT*.json, snap_*.pt (block-bootstrap snapshots written by the diagnostic generator, snap_every), tokens_2d.pt (token windows: runs/R5a/tokens_2d.pt is 1953 x 256 int64 without BOS, SHA-256 23b44fad30303c6a...; runs/R2-bos-mask/tokens_2d.pt has a BOS column; runs/R2-bos-mask/tokens_2d.pt[:, 1:] equals the R5 windows), selfcheck.json, code.sha256, gpu_mem.log, plots. MANIFEST.sha256 lists the SHA-256 of every file.
Association run (runs/phase1/association-full): statistics only (no text, no weights). The sidecar saves the quantized side's own firing counts and activation sums, which the older survival.pt files do not; status definitions and limits are in the repository (reports/association_retention/interpretation.md). It is exploratory on already examined data.
Phase 1 file types: survival.pt, selectivity.pt, per_language.csv, summary.csv, meta.json, gpu_mem.log.
survival.ptkeys:languages,bits,n(tokens per language),corr(per rung, shape[27, 16384]: Pearson correlation between the full-precision and the quantized SAE activations of each feature in each language),firing_rate,mean_activation,max_activation(shape[27, 16384]),loss_sum,loss_n,pooled_corr,pooled_firing_rate.firing_rate,mean_activationandmax_activationdescribe the full-precision features only; the quantized side's firing rates, means and maxima were not saved.corris a correlation of activation patterns: it does not show that a feature keeps its language association.selectivity.ptkeys includerates,mean_activation,S,S_rank,nu,peak,supported,selective,controls,stable_count,strata(tagging on the dev split, full precision).
Status notes
Phase 1 results and their reading are in the repository (reports/phase1_results.md, reports/phase1_summary.md). The pre-registered selective-versus-control comparisons were null in the primary (p1-mask, eq) stream; later analyses are exploratory and correlational (27 languages). The project later restricted itself to the mask regime and no longer compares numerically with Duan (2026); the none and baseline folders are kept as historical records.
Download: hf download karma-skz/always-late-artifacts --repo-type dataset --local-dir artifacts.
- Downloads last month
- 13