Datasets:
activation_vector listlengths 1.54k 1.54k | doc_id stringlengths 12 21 | activation_layer int64 23 23 | domain stringclasses 15
values | labeler_model stringclasses 3
values | detokenized_text_truncated stringlengths 74 4.81k ⌀ | response stringlengths 123 312 | old_label_verbatim stringlengths 66 167 | source_text_recovered stringlengths 74 5.31k | prior_weights listlengths 28 69 | pw_mean float64 0.53 0.89 | pw_fraczero float64 0.07 0.29 | pw_ntok int64 28 69 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
[
1.7578125,
0.0908203125,
0.5703125,
-0.22265625,
-0.1796875,
0.0361328125,
0.1328125,
0.3125,
-1.1640625,
0.3046875,
0.52734375,
-0.333984375,
-0.44921875,
-1.0390625,
0.232421875,
-0.12255859375,
-0.9375,
3.6875,
-0.291015625,
-0.1416015625,
-0.482421875,
-0.0234375,
0.1... | arx_00000006 | 23 | arxiv | opus-4.8 | null | <explanation>Model tracks that comparing 4H and 6H SiC polytypes isolates SPhP monopole-mode coupling to a zone-folded optical phonon, with 'phonon(s)' the expected next word.</explanation> | <explanation>
SiC nanopillar SHG via localized surface phonon-polaritons in reststrahlen band
</explanation> | we report on strong enhancement of mid - infrared second harmonic generation ( shg ) from sic nanopillars due to the resonant excitation of localized surface phonon - polaritons within the reststrahlen band .
the magnitude of the shg peak at the monopole mode experiences a strong dependence on the resonant frequency ... | [
0.9998947978019714,
0.0020738840103149414,
0.00031304359436035156,
0.9999999403953552,
0.9999416470527649,
0.9999589920043945,
0.9999915361404419,
0.996120810508728,
0.9545192122459412,
0.9997556209564209,
0.9274266362190247,
0.15553951263427734,
0.8336800932884216,
0.04460376501083374,
... | 0.776445 | 0.12766 | 47 |
[
-0.53125,
0.9453125,
-0.298828125,
-2.8125,
-0.171875,
0.9375,
-0.2158203125,
0.80078125,
-0.10791015625,
0.380859375,
0.06201171875,
-0.0791015625,
-0.359375,
-0.12255859375,
0.236328125,
0.73046875,
1.1640625,
1.015625,
0.396484375,
-0.72265625,
0.0791015625,
0.66796875,
... | arx_00000011 | 23 | arxiv | opus-4.8 | null | <explanation>Model tracks that the derived @xmath7 branching ratio, built on CLEO-c's @xmath6 measurement, now functions as the normalization anchor for other charm branching-fraction extractions.</explanation> | <explanation>
CLEO-c charm meson absolute branching fractions, Cabibbo-suppressed decays, CP/T violation, mixing limits
</explanation> | i discuss new results on absolute branching ratios of charm mesons into specific exclusive final states , cabibbo suppressed decay rates , inclusive decays to @xmath0 mesons , limits on @xmath1 mixing , cp violation and t violation .
preliminary results from cleo - c now dominate the world average absolute branching ... | [
0.9998947978019714,
0.0020738840103149414,
0.00031304359436035156,
0.9999999403953552,
0.9999416470527649,
0.9999589920043945,
0.4198717474937439,
0.9999834895133972,
0.999971330165863,
0.04576045274734497,
0.09116411209106445,
0.9750592708587646,
0.9999895691871643,
0.9786493182182312,
... | 0.651331 | 0.217391 | 46 |
[
-3.796875,
0.98046875,
0.8359375,
0.77734375,
-0.10546875,
0.091796875,
-0.6015625,
0.0296630859375,
-0.8046875,
1.3046875,
-0.490234375,
-0.054931640625,
-1.0078125,
-0.578125,
0.23046875,
1.40625,
-0.8125,
3.34375,
-0.43359375,
-0.4453125,
-1.40625,
0.486328125,
-0.2773... | arx_00000006 | 23 | arxiv | opus-4.8 | null | <explanation>Model tracks that comparing 4H and 6H SiC polytypes isolates SPhP monopole-mode coupling to a zone-folded optical phonon, with 'phonon(s)' the expected next word.</explanation> | <explanation>
Resonant phonon-polariton excitation enhancing SHG peak at SiC monopole mode
</explanation> | we report on strong enhancement of mid - infrared second harmonic generation ( shg ) from sic nanopillars due to the resonant excitation of localized surface phonon - polaritons within the reststrahlen band .
the magnitude of the shg peak at the monopole mode experiences a strong dependence on the resonant frequency ... | [
0.9998947978019714,
0.0020738840103149414,
0.00031304359436035156,
0.9999999403953552,
0.9999416470527649,
0.9999589920043945,
0.9999915361404419,
0.996120810508728,
0.9545192122459412,
0.9997556209564209,
0.9274266362190247,
0.15553951263427734,
0.8336800932884216,
0.04460376501083374,
... | 0.776445 | 0.12766 | 47 |
[
3.046875,
-0.5703125,
0.1328125,
3.3125,
-0.361328125,
1.078125,
0.490234375,
0.796875,
-1.578125,
-0.25,
0.671875,
0.3671875,
-1.03125,
-0.703125,
0.39453125,
-0.474609375,
-0.02880859375,
1.796875,
-0.0263671875,
-0.375,
-1.0234375,
0.1474609375,
-0.169921875,
-3.125,... | arx_00000003 | 23 | arxiv | opus-4.8 | null | <explanation>Model must complete 'suppress the' with 'scintillations', tracking that large-amplitude, finite-time phase fluctuations specifically weaken (not eliminate) the diffuser's suppression.</explanation> | <explanation>
Random phase diffuser's effect on laser light scintillation index, finite-time temporal phase variations
</explanation> | the effect of a random phase diffuser on fluctuations of laser light ( scintillations ) is studied .
not only spatial but also temporal phase variations introduced by the phase diffuser are analyzed .
the explicit dependence of the scintillation index on finite - time phase variations is obtained for long propagati... | [
0.9998947978019714,
0.0020738840103149414,
0.00031304359436035156,
0.9999999403953552,
0.9999430179595947,
0.938865065574646,
0.9991583228111267,
0.999728262424469,
0.7425354719161987,
0.965944766998291,
0.7362525463104248,
0.8225442171096802,
0.9826129078865051,
0.99980229139328,
0.9777... | 0.747437 | 0.139535 | 43 |
[
1.09375,
0.81640625,
0.69921875,
1.3828125,
-0.44140625,
1.1328125,
-0.8515625,
-0.212890625,
-0.90625,
0.28515625,
0.96484375,
-0.427734375,
-1.0625,
-1.2421875,
0.0198974609375,
-0.08642578125,
0.4140625,
1.875,
0.1708984375,
-1.078125,
0.0791015625,
-0.33203125,
0.4472... | arx_00000002 | 23 | arxiv | opus-4.8 | null | <explanation>Model tracks that the paper sides with Barbi-Salerno's ensemble-based finding of current reversal without bifurcation, and must supply a second enumerated result after 'as well as'.</explanation> | <explanation>
Current reversal in deterministic ratchets without bifurcations, contra Mateos
</explanation> | in 84 , 258 ( 2000 ) , mateos conjectured that current reversal in a classical deterministic ratchet is associated with bifurcations from chaotic to periodic regimes .
this is based on the comparison of the current and the bifurcation diagram as a function of a given parameter for a periodic asymmetric potential .
... | [
0.9998947978019714,
0.0020738840103149414,
0.00031304359436035156,
0.9999999403953552,
0.9999416470527649,
0.9999589920043945,
0.4198717474937439,
0.9999871253967285,
0.9999996423721313,
0.04484212398529053,
0.9999990463256836,
0.9949973821640015,
0.9984695315361023,
0.999823272228241,
0... | 0.775375 | 0.139535 | 43 |
[
1.453125,
-0.88671875,
0.703125,
3.375,
-0.1787109375,
0.71875,
1.0703125,
1.15625,
-1.828125,
0.14453125,
-0.232421875,
0.19140625,
-1.0546875,
0.2412109375,
-0.365234375,
0.474609375,
0.8046875,
1.8125,
0.034423828125,
0.1328125,
-2.71875,
0.79296875,
-0.234375,
0.289... | arx_00000007 | 23 | arxiv | opus-4.8 | null | <explanation>Model must complete 'compact, ultra-fast and low' with 'power', tracking MTJ magnetization dynamics as the established analog to frequency-dependent biological synaptic strengthening.</explanation> | <explanation>
Volatile short-term plasticity in nanoelectronic synapses mimicking biological synaptic memory
</explanation> | synaptic memory is considered to be the main element responsible for learning and cognition in humans .
although traditionally non - volatile long - term plasticity changes have been implemented in nanoelectronic synapses for neuromorphic applications , recent studies in neuroscience have revealed that biological syn... | [
0.9998947978019714,
0.0020738840103149414,
0.00031304359436035156,
0.9999999403953552,
0.9999430179595947,
0.938865065574646,
0.9991583228111267,
0.9999865889549255,
0.9984946846961975,
0.9984456300735474,
0.16656124591827393,
0.9056098461151123,
0.9948815107345581,
0.9639957547187805,
0... | 0.829417 | 0.105263 | 38 |
[
-0.8984375,
1.5703125,
1.3984375,
-0.34375,
-0.3984375,
-1,
-0.82421875,
0.0927734375,
1.0234375,
-0.9921875,
-0.90625,
-0.302734375,
-0.3046875,
0.65625,
0.66796875,
-0.49609375,
1.078125,
2.390625,
0.25,
-0.1123046875,
-2.265625,
0.236328125,
0.427734375,
-0.96484375,... | arx_00000001 | 23 | arxiv | opus-4.8 | null | <explanation>Model is mid-macro completing '@x' into a LaTeX symbol for the decay constant just derived from CLEO-c's tagged sample near the @xmath3 peak energy.</explanation> | <explanation>
CLEO-c leptonic decay channel result and decay-constant determination
</explanation> | we have studied the leptonic decay @xmath0 , via the decay channel @xmath1 , using a sample of tagged @xmath2 decays collected near the @xmath3 peak production energy in @xmath4 collisions with the cleo - c detector .
we obtain @xmath5 and determine the decay constant @x | [
0.9998947978019714,
0.0020738840103149414,
0.00031304359436035156,
0.9999999403953552,
0.9981415867805481,
0.9997918605804443,
0.001032710075378418,
0.9997913241386414,
0.9999949932098389,
0.9999939203262329,
0.6935190558433533,
0.8398904204368591,
0.9964370131492615,
0.7369669079780579,
... | 0.737895 | 0.142857 | 42 |
[
1.1328125,
1.25,
0.0072021484375,
1.140625,
0.232421875,
1.9296875,
-0.271484375,
0.42578125,
0.259765625,
0.427734375,
2.546875,
0.70703125,
0.134765625,
0.42578125,
-0.412109375,
-0.5546875,
-0.70703125,
2.765625,
-0.12255859375,
-0.80859375,
0.23046875,
-0.37109375,
0.... | arx_00000012 | 23 | arxiv | opus-4.8 | null | <explanation>Model is inside LaTeX macro-numbering boilerplate ('#1;#2;#3;#'), tracking sequential placeholder indices rather than physics content, expecting '#4' next rather than any cosmological claim.</explanation> | <explanation>
Open cosmogony parameter constraints from DMR sky-map orthogonal mode analysis
</explanation> | cut - sky orthogonal mode analyses of the @xmath0-dmr 53 and 90 ghz sky maps are used to determine the normalization of a variety of open cosmogonical models based on the cold dark matter scenario . to constrain the allowed cosmological - parameter range for these open cosmogonies , the predictions of the dmr - normali... | [
0.9998947978019714,
0.0020738840103149414,
0.00031304359436035156,
0.9999999403953552,
0.9981415867805481,
0.999980628490448,
0.9999387264251709,
0.996073842048645,
0.9649669528007507,
0.9999902844429016,
0.677894115447998,
0.9981386065483093,
0.9979628324508667,
0.9909225702285767,
0.56... | 0.753099 | 0.139535 | 43 |
[
0.94921875,
-1.1953125,
0.33984375,
-0.341796875,
-0.134765625,
-0.189453125,
0.1552734375,
0.734375,
0.234375,
1.3046875,
0.2392578125,
0.7890625,
-1.25,
-0.578125,
-0.443359375,
0.734375,
-0.259765625,
1.9296875,
0.2109375,
-0.515625,
-0.75390625,
0.28515625,
0.1875,
... | arx_00000008 | 23 | arxiv | opus-4.8 | null | <explanation>Model tracks that b-scale pose encoding enables one-shot model placement rather than search-based optimization, and must complete the closing claim about delineation's resulting ease/accuracy.</explanation> | <explanation>
Ball-scale prior shape models for single-shot 3D object recognition
</explanation> | this paper investigates , using prior shape models and the concept of ball scale ( b - scale ) , ways of automatically recognizing objects in 3d images without performing elaborate searches or optimization .
that is , the goal is to place the model in a single shot close to the right pose ( position , orientation , a... | [
0.9998947978019714,
0.0020738840103149414,
0.00031304359436035156,
0.9999999403953552,
0.9999416470527649,
0.9999589920043945,
0.9999849796295166,
0.9604795575141907,
0.9998825192451477,
0.9999967217445374,
0.999153196811676,
0.9998617172241211,
0.99993497133255,
0.11266648769378662,
0.8... | 0.767389 | 0.170732 | 41 |
[2.421875,-1.640625,-0.578125,1.6796875,-0.1181640625,0.84765625,0.1318359375,0.1328125,-1.84375,-0.(...TRUNCATED) | arx_00000003 | 23 | arxiv | opus-4.8 | null | "<explanation>Model must complete 'suppress the' with 'scintillations', tracking that large-amplitud(...TRUNCATED) | "<explanation>\nRandom phase diffuser effect on laser-light scintillations, spatial and temporal var(...TRUNCATED) | "the effect of a random phase diffuser on fluctuations of laser light ( scintillations ) is studied (...TRUNCATED) | [0.9998947978019714,0.0020738840103149414,0.00031304359436035156,0.9999999403953552,0.99994301795959(...TRUNCATED) | 0.747437 | 0.139535 | 43 |
End of preview. Expand in Data Studio
Gemma-4-E2B layer-23 activation corpus, relabeled (v1)
1356 training rows for an activation verbalizer. Each row pairs a residual-stream
activation captured at layer 23 of google/gemma-4-E2B with a natural-language label
describing what the model must have integrated at that position to predict its next
token. This is the training set behind
Solshine/gemma-4-e2b-nla-L23-av-priordev-relabel-v1-wd3.
Why it exists
An audit of the previous version of this corpus found that all 1356 labels were document topic summaries ("Current reversal in deterministic ratchets without bifurcations") rather than the feature attributions the labeling instruction asked for. The round-trip reconstruction metric used to evaluate verbalizers could not detect that on its own (labels that are unverifiable reconstruct just as well as verifiable ones). Every row was therefore relabeled against a corrected rubric that states the core question, gives four worked bad-to-good examples, and adds a self-check ("could this label describe several different documents in this domain equally well? If yes, revise"). The old label is preserved on every row.
Files
av_sft_balanced_weighted.parquet (single file, 1356 rows, 13 columns):
| Column | Type | Meaning |
|---|---|---|
activation_vector |
float32[1536] | Residual-stream activation at layer 23, at the cut position. |
doc_id |
str | Source-document id (prefix names the domain). Several rows can share a doc_id at different cut positions. |
activation_layer |
int | 23. |
domain |
str | One of 15: arxiv, deception, fineweb, global_opinions, legal, math, medicine, mmlu_moral, news, news_rl, persuasion, pku_safety, reviews, science, wiki (136 rows each for six of them, 53–109 for the rest). |
labeler_model |
str | Which pipeline produced the original label (703 rows opus-4.8, 653 rows two Gemini pipelines). The new label was written by a single relabeling pass for every row. |
detokenized_text_truncated |
str | The source text up to the cut, as shipped in the original corpus (null on the 703 opus-4.8 rows). |
source_text_recovered |
str | The source text up to the cut for every row, recovered by joining doc_id back to the raw extraction. Non-null everywhere. |
response |
str | The new feature-attribution label, wrapped in <explanation>...</explanation>. |
old_label_verbatim |
str | The original topic-summary label, unchanged. |
prior_weights |
float32[seq] | Per-token loss weights for prior-deviation training: tokens the un-conditioned model already predicts well are down-weighted. Recomputed for the new labels. |
pw_mean, pw_fraczero, pw_ntok |
float, float, int | Summary statistics of prior_weights for the row (mean 0.753 and fraction near zero 0.142 across the corpus). |
Quality
An independent audit of a random 120-row sample against the same rubric returned 99 pass, 17 borderline, 4 fail (82.5% clean; Wilson 95% interval roughly 0.75 to 0.88), up from 0 of 1424 under the original labels. The audit named four failure modes, all present at low rates in this file:
- wrapper-only relabels: "Model tracks ..." prepended to what is still a topic summary;
- cut-position inaccuracy: a label describing as "about to be predicted" something the source text already stated before the cut, or the reverse;
- one confirmed case of the same label reused for two different cut positions of one document;
- one confirmed case of two rows for the same document giving contradictory accounts of where the cut falls.
Rows were not filtered on audit outcome; the file is the corpus as trained on.
Relationship to the original corpus
This is not the original 1757-row training corpus with new labels. It is a 1356-row balanced subset drawn earlier in the project, sharing about 81% of its documents with the original. Experiments that compare adapters trained on this file against adapters trained on the original corpus therefore isolate "which corpus", not "label style alone". A row-matched relabel of the original corpus is planned.
Provenance
Source documents come from public corpora (FineWeb-Edu, Wikipedia, arXiv abstracts,
legal opinions, PubMed, MMLU, PKU-SafeRLHF, and project-internal review, persuasion,
and deception-scenario text). Activations were extracted with google/gemma-4-E2B
under 4-bit NF4 quantisation on a 4 GB GPU. Labels were written by a commodity API
model from the recovered source text and audited by a different model.
License
CC BY 4.0 for the labels and metadata. Source-text excerpts retain their original licenses; the corpus is distributed for interpretability research.
- Downloads last month
- 44