Dataset Viewer
Auto-converted to Parquet Duplicate
example_idx
int32
0
1.32k
is_correct
bool
2 classes
gold_answer
stringlengths
48
1.07k
rollout_idx
int32
0
39
sentence_idx
int32
0
10
sentence_text
stringlengths
11
1.43k
sentence_embedding
listlengths
384
384
cos_sim_to_full_nla
float32
0.01
1
critic_cos_sim
float32
-0.2
0.94
critic_mse
float32
0.13
2.4
contains_gold_answer
bool
2 classes
contains_pred_answer
bool
2 classes
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
0
0
Structured math problem format with labeled steps showing a calculation involving sets and combinations, resolving to a final answer about a combined quantity of objects.
[ 0.055825624614953995, 0.11256259679794312, -0.02502628229558468, -0.01834890805184841, -0.0926840752363205, 0.03180643916130066, 0.005110098514705896, 0.06325977295637131, -0.03751588985323906, 0.025227004662156105, -0.08662672340869904, -0.0918540209531784, 0.04169633239507675, 0.05047246...
0.678577
0.650093
0.699815
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
0
1
The sentence "Thus, the sum is 8+2=10.
[ -0.03528999164700508, 0.10368276387453079, 0.0734054446220398, 0.055595189332962036, 0.020186826586723328, 0.0610717236995697, 0.036680448800325394, 0.10543064773082733, -0.00017384921375196427, -0.014249595813453197, 0.042238350957632065, -0.039575982838869095, -0.051424477249383926, -0.0...
0.641102
0.68647
0.62706
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
0
2
Answer: " mirrors a concluding display format, strongly implying the final answer "4" is formatted as "S 3" or "The total number." Final token " " opens a numeric expression completion ("答案 3 "), directly continuing the answer statement "Result: 3 +", immediately requiring "10" or "3" to close the formatted result,...
[ 0.01265402790158987, 0.06456294655799866, 0.020901313051581383, 0.005836859345436096, -0.016757849603891373, 0.06814084947109222, 0.02379561774432659, 0.07620761543512344, 0.1136503666639328, 0.02155439555644989, 0.032927799969911575, -0.07606571167707443, -0.01509094424545765, 0.040307726...
0.871148
0.834609
0.330782
true
true
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
1
0
Structured math format with numbered steps showing a combined expression, interpreting a word problem with parentheses and sum notation.
[ -0.038378581404685974, 0.13881918787956238, 0.0356498658657074, -0.01045500859618187, -0.09329802542924881, 0.038551121950149536, -0.02895232103765011, 0.07400038838386536, -0.015290085226297379, 0.009130523540079594, -0.018845433369278908, -0.07430411130189896, 0.040453314781188965, 0.046...
0.650825
0.586527
0.826947
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
1
1
The phrase "Thus, the final answer is: 4 + 2 + 1 = " signals the conclusion, confirming the calculated total value "5," and the final answer format "Box 3" implies a formatted answer display completing the box sum concept.
[ 0.022020455449819565, 0.10446926951408386, 0.018041862174868584, 0.01751212775707245, -0.009935908019542694, 0.02502295933663845, 0.013718797825276852, -0.0035490524023771286, 0.07101836055517197, 0.05307778716087341, -0.05627935007214546, -0.09119988232851028, -0.05583158880472183, -0.008...
0.854002
0.712264
0.575472
true
true
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
1
2
Final token " " is mid-number in "答案 " (answer 3), part of a concluding summary statement "答案: 5" — immediately expects "5" or "8" or "boxed" to close the arithmetic result, completing the box-sum or total-value conclusion.
[ 0.00827029813081026, 0.05729145184159279, 0.13012182712554932, 0.02184094488620758, 0.04462118446826935, 0.003507959423586726, 0.04104461893439293, 0.042145054787397385, 0.09535621851682663, 0.043852392584085464, 0.033436354249715805, -0.06488659977912903, -0.0031131524592638016, 0.0109821...
0.801073
0.848509
0.302983
true
true
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
2
0
Structured math format with numbered questions and calculations showing combined terms from a set, presenting a concise summary result.
[ 0.027292905375361443, 0.1295594871044159, -0.020511101931333542, -0.04931287094950676, -0.10139521956443787, 0.05524592101573944, -0.004044088535010815, 0.08026813715696335, -0.01286064088344574, 0.01732456497848034, -0.048969075083732605, -0.0817866399884224, 0.04319427162408829, 0.054442...
0.615002
0.629166
0.741669
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
2
1
The phrase "Therefore, the total number of items is 7+3= " completes the answer, following a concluding phrase "The final answer: 10" — strongly implying the formatted answer "4" or "the box" closes the sum result.
[ 0.0025493318680673838, 0.09480302780866623, 0.03618471696972847, 0.07312679290771484, -0.011389855295419693, 0.05038969963788986, 0.06118595227599144, 0.027068424969911575, 0.05598586052656174, 0.08464603126049042, -0.02926986664533615, -0.05972513556480408, -0.061765193939208984, -0.01253...
0.808255
0.861225
0.277551
true
true
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
2
2
Final token " " is mid-expression ("答案 "), part of a concluding summary line "答案:4.
[ 0.013838798739016056, 0.07193610817193985, 0.15241184830665588, 0.008767144754529, 0.08495813608169556, -0.0012306120479479432, 0.02636539936065674, 0.06027096137404442, 0.07154678553342819, 0.017002370208501816, 0.06470487266778946, -0.09372145682573318, 0.05267925187945366, 0.02854843251...
0.621545
0.768698
0.462604
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
2
3
" — immediately expects "4" or "10" or "boxed" or "sum" to complete the numeric answer describing the combined box count.
[ 0.06754162907600403, 0.023443792015314102, -0.002427878323942423, 0.00886360090225935, -0.0724560096859932, 0.0872529149055481, 0.00609436072409153, 0.028811169788241386, 0.01444940734654665, -0.031584471464157104, 0.005968519952148199, -0.0918523445725441, 0.0015375104267150164, 0.0155077...
0.695842
0.767051
0.465898
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
3
0
Structured math problem format with numbered steps and explanations, showing calculations for a combined set describing a geometric shape's properties.
[ -0.0024980478920042515, 0.11476132273674011, 0.00023101600527297705, -0.05962895229458809, -0.10541142523288727, -0.010571143589913845, -0.027609558776021004, 0.054557256400585175, -0.06093967333436012, -0.002571668243035674, -0.07577827572822571, -0.09853821992874146, 0.07359441369771957, ...
0.608654
0.577963
0.844075
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
3
1
The phrase "Therefore, the answer is: 5 + 3 + 2 = 10" signals the final answer format, completing the "The total number of items" summary, strongly implying the closed form "8" or "A" as the box's answer.
[ -0.02800191193819046, 0.1021370142698288, 0.07024087011814117, 0.03676588460803032, 0.025605887174606323, 0.061459075659513474, 0.04018082469701767, 0.05373422056436539, 0.0529206357896328, 0.08477923274040222, 0.010850024409592152, -0.04835285618901253, -0.06203470006585121, -0.0168101489...
0.837614
0.675668
0.648665
true
true
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
3
2
Final token " " is mid-expression (" " 8"), part of the concluding answer display ("Answer: "), immediately expecting "8" or "10" to close the numeric answer, continuing the box-sum notation pattern.
[ 0.009558102115988731, 0.12200126051902771, 0.09024901688098907, 0.03497907146811485, 0.005301621276885271, 0.07850479334592819, 0.022967083379626274, 0.10048996657133102, 0.06781710684299469, 0.023958448320627213, -0.03539947792887688, -0.03925315663218498, -0.04475374519824982, 0.00870088...
0.798272
0.847082
0.305836
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
4
0
Structured math format with numbered steps and labeled expressions showing a combined quantity problem, presenting a worked solution about a geometric shape's total count.
[ 0.022430213168263435, 0.06761694699525833, -0.01261131465435028, -0.035945214331150055, -0.12321265041828156, 0.06330648809671402, -0.027517050504684448, 0.06516119092702866, 0.009444531984627247, -0.008120912127196789, -0.03524308651685715, -0.0766223892569542, 0.05652891471982002, 0.1116...
0.697454
0.606643
0.786714
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
4
1
The sentence "Thus, the final answer is: 8+2=10" signals the answer format "Answer: 5" is being completed, with the concluding phrase "The sum is " implying the box format result.
[ 0.004786393139511347, 0.12577134370803833, 0.07128492742776871, 0.036678433418273926, 0.009957749396562576, 0.07254316657781601, -0.009764720685780048, 0.029418425634503365, 0.02581714279949665, 0.017893437296152115, -0.036902762949466705, -0.028879664838314056, -0.07714463025331497, 0.005...
0.859207
0.838673
0.322655
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
4
2
Final token " " is mid-expression (" 5 "), part of the answer confirmation statement "Answer: " structure, immediately requiring the numeric value "10" to close the box total, likely followed by "10" or "sum" or "5" completing the summary description.
[ -0.022012392058968544, 0.08019884675741196, 0.05824729800224304, 0.03409822657704353, 0.053151149302721024, 0.031826552003622055, 0.07637817412614822, 0.05522201210260391, 0.07617593556642532, 0.04824851453304291, 0.0030931655783206224, -0.07518258690834045, -0.04191436991095543, 0.0071905...
0.812361
0.85362
0.292759
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
5
0
Structured math problem format with numbered steps showing calculations, defining terms like "sum," "product," and combining the two values from a set.
[ -0.009488765150308609, 0.08608533442020416, -0.024337027221918106, -0.07614092528820038, -0.12178044766187668, 0.03823802247643471, -0.009034351445734501, 0.06853603571653366, -0.06120491027832031, 0.0038883842062205076, -0.0517197921872139, -0.055829066783189774, 0.07686518132686615, 0.03...
0.701
0.618696
0.762609
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
5
1
The sentence "Thus, the final answer is: 8 + 2 = 10" concludes with a formatted answer display, strongly implying the final answer expression "4" or "10" is being confirmed or summarized.
[ -0.013236635364592075, 0.1349935382604599, 0.060677845031023026, 0.03670806437730789, 0.024053646251559258, 0.06186061352491379, -0.01240402739495039, 0.11628041416406631, 0.047179318964481354, 0.014577196910977364, 0.03364880010485649, -0.04517580196261406, -0.05308845639228821, 0.0445133...
0.784658
0.768293
0.463414
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
5
2
Final token " " is mid-expression in "答案 3 " — part of a concluding summary statement "答案: 4" structure, immediately expecting "10" or "4" to close the answer, likely followed by "5" or "the combined set" confirming the total object.
[ 0.022642623633146286, 0.01782943308353424, 0.11153139173984528, 0.022889750078320503, 0.05203932151198387, 0.002472447697073221, 0.04503500461578369, 0.028864530846476555, 0.11205839365720749, 0.03559959679841995, 0.04211418330669403, -0.12183232605457306, 0.004279696848243475, 0.011461619...
0.770714
0.807703
0.384593
true
true
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
6
0
Structured math problem format with step-by-step solutions presenting a word problem involving sets and quantities, following a pattern of combining values to find a total count.
[ 0.04964851588010788, 0.08634060621261597, 0.0174043420702219, -0.01362666953355074, -0.13829182088375092, 0.09740294516086578, 0.012238760478794575, 0.07220156490802765, -0.035870060324668884, 0.02026721090078354, -0.017268283292651176, -0.08967314660549164, 0.11190365254878998, 0.06156509...
0.703598
0.595469
0.809063
false
false
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
6
1
The answer statement "Thus, 3 + 8 + 5 = 16" concludes with "The final answer is: " signaling the final formatted answer, confirming the combined set notation result.
[ 0.030473290011286736, 0.10472665727138519, 0.056773699820041656, 0.013008197769522667, -0.023827068507671356, 0.012636053375899792, 0.028343815356492996, 0.04618777334690094, 0.010604380629956722, 0.0068086679093539715, 0.008639739826321602, -0.07198890298604965, -0.0014947816962376237, -0...
0.820108
0.840864
0.318273
true
true
1
true
It takes 2/2=<<2/2=1>>1 bolt of white fiber So the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric #### 3
6
2
Final token " " opens the numeric answer "3 " — part of the final answer display ("Answer: S "), strongly expecting "16" or "3" to close the numeric value, completing the concise summary phrase "The total is 16".
[ 0.04000880941748619, 0.07713738083839417, 0.017068222165107727, -0.008833191357553005, -0.004749999847263098, 0.034991465508937836, -0.004210472106933594, 0.08630675077438354, 0.08849970996379852, 0.0652318000793457, -0.003238566219806671, -0.11491045355796814, -0.03695478290319443, 0.0397...
0.781787
0.845945
0.30811
true
true
End of preview. Expand in Data Studio

nla-thought-anchors-answer-rollouts

Sentence-level NLA rollout dataset — activations extracted at the answer digit token.

For each of 1319 GSM8K test examples, 40 NLA descriptions were sampled from kitft/nla-qwen2.5-7b-L20-av (temperature=0.7) using the layer-20 residual stream activation at the first digit after #### (answer now in context). Each description was split into sentences and scored with kitft/nla-qwen2.5-7b-L20-ar.

See Realmbird/nla-thought-anchors-hash-rollouts for the ####-token variant.

Stats

Examples 1319
Rollouts per example 40
Total sentence rows 162,316
Mean sentences per rollout ~3.1

Schema

Column Type Description
example_idx int32 Index in GSM8K test split
is_correct bool Whether model answer matched gold
gold_answer string Ground-truth answer
rollout_idx int32 0–39
sentence_idx int32 Sentence position within this rollout
sentence_text string The sentence text
sentence_embedding list<float32> len=384 all-MiniLM-L6-v2 embedding
cos_sim_to_full_nla float32 Cos-sim of sentence to its own rollout's full description
critic_cos_sim float32 NLA critic reconstruction cos-sim vs original activation
critic_mse float32 NLA critic reconstruction MSE vs original activation
contains_gold_answer bool Sentence contains the gold answer number
contains_pred_answer bool Sentence contains the predicted answer number

Extraction position comparison

Position Mean critic cos-sim (correct) Mean critic cos-sim (incorrect)
#### token (hash variant) 0.604 0.416
Answer digit token (this dataset) TBD TBD

The answer token has the numeric answer already written into the sequence, making it the most direct encoding of the answer value.

Computing anchor scores

from datasets import load_dataset
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity
import pandas as pd

df = load_dataset("Realmbird/nla-thought-anchors-answer-rollouts", split="train").to_pandas()
embs = np.stack(df["sentence_embedding"].values)

anchor_scores = np.zeros(len(df))
for ex_idx, group in df.groupby("example_idx"):
    idx = group.index
    centroid = embs[idx].mean(axis=0)
    centroid /= np.linalg.norm(centroid) + 1e-12
    anchor_scores[idx] = cosine_similarity(embs[idx], centroid.reshape(1,-1)).flatten()

df["anchor_score"] = anchor_scores

Related

  • Hash-token variant: Realmbird/nla-thought-anchors-hash-rollouts
  • Step 1 activations: Realmbird/nla-thought-anchors-answer
  • Step 2 NLA descriptions: Realmbird/nla-thought-anchors-answer-step2
  • NLA actor: kitft/nla-qwen2.5-7b-L20-av
  • NLA critic: kitft/nla-qwen2.5-7b-L20-ar
Downloads last month
9

Collections including Realmbird/nla-thought-anchors-answer-rollouts