Hugging Face's logo

Datasets:
connections-dev
/
create-interpret-task-general-anchors

Modalities:
Tabular
Text
Formats:
json
Size:
100K - 1M
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community
Dataset Viewer
Auto-converted to Parquet Duplicate
dataset
stringclasses
1 value
instance_id
stringclasses
10 values
sentence_idx
int64
0
189
sentence
stringlengths
10
201
tags
listlengths
1
3
rel_pos
float64
0
0.99
n_sentences
int64
65
191
status
stringclasses
1 value
n_ok
int64
100
100
n_failed
int64
0
0
n_no_answer
int64
0
0
counts
unknown
top_answer
stringclasses
23 values
importance
float64
0
5.46
alpha
float64
0.1
0.1
rat
rat-0004
0
Okay, let's see.
[ "plan_generation" ]
0
83
ok
100
0
0
{ "watch": 84, "light": 4, "clock": 9, "band": 1, "time": 2 }
watch
0.102183
0.1
rat
rat-0004
1
I need to find a fourth word that's related to all three: night, wrist, and stop.
[ "problem_setup" ]
0.012195
83
ok
100
0
0
{ "clock": 20, "watch": 76, "light": 1, "time": 1, "rest": 1, "hand": 1 }
watch
0.136649
0.1
rat
rat-0004
2
Hmm. Let me think about each word and possible connections.
[ "plan_generation" ]
0.02439
83
ok
100
0
0
{ "watch": 87, "band": 1, "clock": 7, "time": 3, "light": 2 }
watch
0.019759
0.1
rat
rat-0004
3
First, night.
[ "fact_retrieval" ]
0.036585
83
ok
100
0
0
{ "watch": 91, "clock": 5, "light": 2, "time": 2 }
watch
0.028234
0.1
rat
rat-0004
4
Words related to night might be dark, evening, sleep, moon, stars, etc.
[ "fact_retrieval" ]
0.04878
83
ok
100
0
0
{ "watch": 85, "clock": 9, "time": 5, "light": 1 }
watch
0.050504
0.1
rat
rat-0004
5
Then wrist—maybe parts of the body, like arm, hand, maybe something related to movement or jewelry?
[ "fact_retrieval" ]
0.060976
83
ok
100
0
0
{ "watch": 90, "clock": 7, "time": 1, "signal": 1, "light": 1 }
watch
0.17376
0.1
rat
rat-0004
6
And stop—like halt, pause, end, maybe a traffic sign, or a verb.
[ "fact_retrieval" ]
0.073171
83
ok
100
0
0
{ "watch": 69, "clock": 23, "hour": 1, "rest": 2, "hand": 2, "light": 1, "time": 2 }
watch
0.063773
0.1
rat
rat-0004
7
Wait, maybe looking for a word that can connect all three in some way.
[ "plan_generation" ]
0.085366
83
ok
100
0
0
{ "watch": 68, "clock": 24, "light": 3, "signal": 1, "hour": 1, "time": 2, "hand": 1 }
watch
0.100143
0.1
rat
rat-0004
8
Let me think of synonyms or categories.
[ "plan_generation" ]
0.097561
83
ok
100
0
0
{ "clock": 16, "watch": 71, "time": 8, "rest": 2, "light": 2, "hand": 1 }
watch
0.059237
0.1
rat
rat-0004
9
Night is a time, wrist is a body part, stop is an action.
[ "fact_retrieval" ]
0.109756
83
ok
100
0
0
{ "time": 5, "watch": 77, "hand": 1, "clock": 10, "rest": 3, "hour": 1, "light": 1, "signal": 2 }
watch
0.117895
0.1
rat
rat-0004
10
Maybe something like "hour"?
[ "plan_generation" ]
0.121951
83
ok
100
0
0
{ "watch": 79, "clock": 15, "light": 2, "time": 2, "band": 1, "rest": 1 }
watch
0.064834
0.1
rat
rat-0004
11
No, not really.
[ "uncertainty_management" ]
0.134146
83
ok
100
0
0
{ "clock": 27, "watch": 65, "light": 3, "band": 1, "rest": 2, "hand": 1, "time": 1 }
watch
0.137166
0.1
rat
rat-0004
12
Or "time"—but how does that connect to wrist?
[ "fact_retrieval", "uncertainty_management" ]
0.146341
83
ok
100
0
0
{ "watch": 72, "clock": 20, "time": 3, "rest": 5 }
watch
0.058643
0.1
rat
rat-0004
13
Maybe not.
[ "uncertainty_management" ]
0.158537
83
ok
100
0
0
{ "watch": 82, "clock": 16, "time": 1, "rest": 1 }
watch
0.026875
0.1
rat
rat-0004
14
Alternatively, think of words that can be parts of other words.
[ "plan_generation" ]
0.170732
83
ok
100
0
0
{ "watch": 84, "rest": 3, "clock": 13 }
watch
0.055062
0.1
rat
rat-0004
15
Like "night" and "wrist" both have parts that could combine with something else.
[ "fact_retrieval" ]
0.182927
83
ok
100
0
0
{ "watch": 77, "clock": 19, "time": 3, "rest": 1 }
watch
0.063548
0.1
rat
rat-0004
16
Wait, maybe "watch"?
[ "plan_generation" ]
0.195122
83
ok
100
0
0
{ "watch": 77, "clock": 16, "light": 4, "band": 1, "time": 1, "rest": 1 }
watch
0.3609
0.1
rat
rat-0004
17
Because a watch is related to time (night), it's worn on the wrist, and a watch can be used to stop time?
[ "active_computation" ]
0.207317
83
ok
100
0
0
{ "watch": 98, "clock": 2 }
watch
0.089784
0.1
rat
rat-0004
18
Or maybe "stop" as in stopping the watch?
[ "active_computation" ]
0.219512
83
ok
100
0
0
{ "watch": 86, "clock": 14 }
watch
0.019629
0.1
rat
rat-0004
19
But is there a word that connects all three?
[ "uncertainty_management" ]
0.231707
83
ok
100
0
0
{ "clock": 19, "watch": 80, "light": 1 }
watch
0.036024
0.1
rat
rat-0004
20
Wait, "wrist" and "stop" could both be parts of a watch.
[ "active_computation" ]
0.243902
83
ok
100
0
0
{ "watch": 71, "clock": 26, "time": 2, "light": 1 }
watch
0.068162
0.1
rat
rat-0004
21
A watch is on the wrist and has a stop function.
[ "active_computation" ]
0.256098
83
ok
100
0
0
{ "watch": 80, "clock": 20 }
watch
0.172256
0.1
rat
rat-0004
22
And night—maybe a watch is used at night.
[ "active_computation" ]
0.268293
83
ok
100
0
0
{ "watch": 96, "clock": 4 }
watch
0.073334
0.1
rat
rat-0004
23
But is there a single word that connects all three?
[ "uncertainty_management" ]
0.280488
83
ok
100
0
0
{ "watch": 84, "clock": 16 }
watch
0.013595
0.1
rat
rat-0004
24
Maybe "watch" is the answer?
[ "result_consolidation" ]
0.292683
83
ok
100
0
0
{ "clock": 19, "watch": 80, "hand": 1 }
watch
0.843596
0.1
rat
rat-0004
25
Let me check again.
[ "self_checking" ]
0.304878
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
26
Night: people use watches at night.
[ "fact_retrieval", "self_checking" ]
0.317073
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
27
Wrist: watches are worn on the wrist.
[ "fact_retrieval", "self_checking" ]
0.329268
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
28
Stop: watches have a stop function to stop the time.
[ "fact_retrieval", "self_checking" ]
0.341463
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
29
So "watch" connects all three.
[ "result_consolidation" ]
0.353659
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
30
That seems plausible.
[ "uncertainty_management" ]
0.365854
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
31
Are there other possibilities?
[ "plan_generation" ]
0.378049
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
32
Another angle: maybe "moon"?
[ "fact_retrieval" ]
0.390244
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
33
Night is associated with the moon, but how does that connect to wrist and stop?
[ "uncertainty_management" ]
0.402439
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
34
Not sure. Or "darkness"—but again, not sure about wrist and stop.
[ "uncertainty_management" ]
0.414634
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
35
Maybe "pause"?
[ "plan_generation" ]
0.426829
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
36
But pause is related to stop, but how to connect with night and wrist?
[ "uncertainty_management" ]
0.439024
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
37
Maybe not.
[ "uncertainty_management" ]
0.45122
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
38
Alternatively, think of "hour" again.
[ "plan_generation" ]
0.463415
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
39
Night is divided into hours, but how does that link to wrist and stop?
[ "uncertainty_management" ]
0.47561
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
40
Not sure. Wait, maybe "clock"?
[ "plan_generation" ]
0.487805
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
41
Similar to watch.
[ "fact_retrieval" ]
0.5
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
42
Clocks can be stopped, are related to time (night), and can be on the wrist (like a wristwatch).
[ "active_computation" ]
0.512195
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
43
So "clock" might also work.
[ "result_consolidation" ]
0.52439
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
44
But the answer might be "watch" or "clock".
[ "uncertainty_management" ]
0.536585
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
45
Which is more likely?
[ "uncertainty_management" ]
0.54878
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
46
The question says a fourth word related to all three.
[ "problem_setup" ]
0.560976
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
47
Let me think again.
[ "plan_generation" ]
0.573171
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
48
If the word is "watch", then yes.
[ "self_checking" ]
0.585366
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
49
If "clock", same.
[ "self_checking" ]
0.597561
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
50
But maybe there's another word.
[ "plan_generation" ]
0.609756
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
51
Wait, what about "pause"?
[ "plan_generation", "fact_retrieval" ]
0.621951
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
52
No. "Stop" is a verb.
[ "self_checking", "uncertainty_management" ]
0.634146
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
53
Maybe "halt"?
[ "plan_generation" ]
0.646341
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
54
Not sure. Alternatively, think of parts of the wrist.
[ "uncertainty_management", "plan_generation" ]
0.658537
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
55
Like "vein" or "bone", but how does that connect to night and stop?
[ "fact_retrieval", "uncertainty_management" ]
0.670732
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
56
Maybe not.
[ "uncertainty_management" ]
0.682927
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
57
Another approach: think of a word that is a part of all three.
[ "plan_generation" ]
0.695122
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
58
For example, "night" and "wrist" both have "ight" and "rist" but that's not helpful.
[ "fact_retrieval", "uncertainty_management" ]
0.707317
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
59
Wait, "night" can be a noun, "stop" a verb, "wrist" a noun.
[ "fact_retrieval" ]
0.719512
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
60
Maybe a word that can function as all three?
[ "plan_generation" ]
0.731707
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
61
Not sure. Alternatively, think of a word that is a synonym for all three.
[ "uncertainty_management", "plan_generation" ]
0.743902
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
62
But "night", "wrist", "stop" are different categories.
[ "fact_retrieval", "uncertainty_management" ]
0.756098
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
63
Wait, maybe "hand"?
[ "plan_generation", "fact_retrieval" ]
0.768293
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
64
Night is related to a clock's hand, wrist is part of the hand, and "stop" could relate to stopping the hand of a clock.
[ "active_computation" ]
0.780488
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
65
But "hand" might be the word.
[ "result_consolidation" ]
0.792683
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
66
Let me check.
[ "plan_generation" ]
0.804878
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
67
Hand is part of the wrist, but how does it connect to night?
[ "self_checking", "uncertainty_management" ]
0.817073
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
68
A clock's hand is part of a clock, which is related to time (night).
[ "fact_retrieval", "active_computation" ]
0.829268
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
69
And stopping the hand.
[ "active_computation" ]
0.841463
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
70
So maybe "hand" is the word.
[ "result_consolidation" ]
0.853659
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
71
But then again, "watch" is more directly connected.
[ "uncertainty_management" ]
0.865854
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
72
So which is more accurate?
[ "uncertainty_management" ]
0.878049
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
73
Alternatively, think of "time"—but how does that connect to wrist and stop?
[ "plan_generation", "fact_retrieval" ]
0.890244
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
74
Maybe not.
[ "uncertainty_management" ]
0.902439
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
75
Wait, the answer might be "watch" because it's the most direct.
[ "result_consolidation" ]
0.914634
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
76
Let me confirm.
[ "plan_generation" ]
0.926829
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
77
Night: you can use a watch during the night.
[ "fact_retrieval" ]
0.939024
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
78
Wrist: a watch is worn on the wrist.
[ "fact_retrieval" ]
0.95122
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
79
Stop: you can stop a watch.
[ "fact_retrieval" ]
0.963415
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
80
So yes, "watch" connects all three.
[ "result_consolidation" ]
0.97561
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0004
81
That seems to fit.
[ "result_consolidation" ]
0.987805
83
ok
100
0
0
{ "watch": 100 }
watch
0
0.1
rat
rat-0006
0
Okay, let's see.
[ "plan_generation" ]
0
93
ok
100
0
0
{ "bicycle": 2, "motion": 48, "swing": 5, "chair": 36, "seat": 3, "balance": 2, "ride": 2, "top": 1, "spin": 1 }
motion
0.165288
0.1
rat
rat-0006
1
I need to find a fourth word that's related to all three: rocking, wheel, and high.
[ "problem_setup" ]
0.01087
93
ok
100
0
0
{ "chair": 43, "motion": 50, "swing": 4, "top": 1, "bicycle": 1, "spin": 1 }
motion
0.128232
0.1
rat
rat-0006
2
Hmm. Let me think about each word and possible connections.
[ "plan_generation" ]
0.021739
93
ok
100
0
0
{ "motion": 41, "swing": 12, "chair": 40, "balance": 2, "bicycle": 2, "spinning": 1, "seat": 1, "ride": 1 }
motion
0.21701
0.1
rat
rat-0006
3
First, "rocking" could refer to the motion of a rocking chair, or maybe something like a rock band?
[ "fact_retrieval" ]
0.032609
93
ok
100
0
0
{ "chair": 48, "seat": 5, "motion": 36, "swing": 6, "swinging": 1, "top": 2, "roll": 1, "cycle": 1 }
chair
0.133976
0.1
rat
rat-0006
4
But how does that connect to a wheel and high?
[ "uncertainty_management" ]
0.043478
93
ok
100
0
0
{ "motion": 32, "chair": 55, "bicycle": 3, "rolling": 1, "balance": 1, "top": 2, "swing": 3, "seat": 2, "roller": 1 }
chair
0.110448
0.1
rat
rat-0006
5
Maybe "rocking" as in a movement.
[ "fact_retrieval" ]
0.054348
93
ok
100
0
0
{ "motion": 40, "cycle": 1, "chair": 48, "balance": 2, "bicycle": 4, "seat": 1, "swing": 3, "movement": 1 }
chair
0.040745
0.1
rat
rat-0006
6
A wheel is a circular object that rotates.
[ "fact_retrieval" ]
0.065217
93
ok
100
0
0
{ "bicycle": 3, "motion": 46, "chair": 45, "swing": 3, "balance": 1, "seat": 2 }
motion
0.101152
0.1
rat
rat-0006
7
High could be a height or a level.
[ "fact_retrieval" ]
0.076087
93
ok
100
0
0
{ "chair": 53, "motion": 35, "spin": 1, "seat": 1, "sway": 1, "cycle": 2, "bicycle": 2, "swing": 1, "balance": 2, "bike": 1, "top": 1 }
chair
0.131473
0.1
rat
rat-0006
8
Wait, maybe there's a common theme.
[ "plan_generation" ]
0.086957
93
ok
100
0
0
{ "motion": 39, "chair": 45, "swing": 3, "bicycle": 4, "balance": 1, "seat": 5, "ride": 1, "cycle": 1, "roll": 1 }
chair
0.240685
0.1
rat
rat-0006
9
Let me think of words that can relate to all three.
[ "plan_generation" ]
0.097826
93
ok
100
0
0
{ "chair": 57, "motion": 37, "movement": 1, "bicycle": 2, "balance": 1, "swing": 1, "bike": 1 }
chair
0.148201
0.1
rat
rat-0006
10
Let's consider synonyms or related concepts.
[ "plan_generation" ]
0.108696
93
ok
100
0
0
{ "chair": 44, "motion": 42, "top": 1, "bicycle": 5, "seat": 3, "ride": 2, "swing": 2, "cycle": 1 }
chair
0.033671
0.1
rat
rat-0006
11
For example, "high" can mean elevated, like a high place.
[ "fact_retrieval" ]
0.119565
93
ok
100
0
0
{ "motion": 42, "chair": 49, "bicycle": 5, "swing": 1, "ride": 1, "seat": 1, "top": 1 }
chair
0.082906
0.1
rat
rat-0006
12
A wheel is part of a vehicle, which might rock if it's moving over rough terrain.
[ "fact_retrieval", "active_computation" ]
0.130435
93
ok
100
0
0
{ "chair": 47, "motion": 38, "cycle": 1, "ride": 2, "swing": 5, "bicycle": 3, "sway": 1, "balance": 1, "seat": 2 }
chair
0.07823
0.1
rat
rat-0006
13
But how does that tie in with "high"?
[ "uncertainty_management" ]
0.141304
93
ok
100
0
0
{ "cycle": 3, "motion": 33, "bicycle": 5, "top": 3, "chair": 45, "seat": 3, "ride": 1, "swing": 4, "balance": 3 }
chair
0.2521
0.1
rat
rat-0006
14
Maybe "high" as in high speed?
[ "fact_retrieval" ]
0.152174
93
ok
100
0
0
{ "motion": 56, "chair": 32, "seat": 3, "bicycle": 3, "swing": 3, "bike": 1, "ride": 1, "cycle": 1 }
motion
0.155427
0.1
rat
rat-0006
15
Like a car going fast, but that's stretching it.
[ "uncertainty_management" ]
0.163043
93
ok
100
0
0
{ "swing": 7, "bicycle": 2, "motion": 42, "chair": 40, "bike": 1, "top": 4, "balance": 3, "seat": 1 }
motion
0.190382
0.1
rat
rat-0006
16
Another angle: maybe the word "wheel" is part of a vehicle, and "rocking" could be the motion of a car on a bumpy road.
[ "active_computation" ]
0.173913
93
ok
100
0
0
{ "motion": 54, "chair": 33, "bicycle": 3, "roll": 1, "spinning": 1, "swing": 5, "seat": 1, "spin": 1, "balance": 1 }
motion
0.108356
0.1
rat
rat-0006
17
But again, how does "high" fit in?
[ "uncertainty_management" ]
0.184783
93
ok
100
0
0
{ "swing": 8, "motion": 48, "bicycle": 8, "chair": 30, "movement": 1, "seat": 2, "top": 1, "ride": 1, "balance": 1 }
motion
0.08324
0.1
End of preview. Expand in Data Studio

simpleqa_verified

Rollouts: 102400 rows written (101900 successful after retries, 0 still failed). finish_reason counts: {'stop': 101897, 'length': 3}. Positions: 1019 sentences over 10 instances, 0 refused. Importance: 1009 rows, 1009 scored, alpha=0.1, mean 0.1947.

Method: Thought Anchors resampling importance. For sentence i, the set is n_rollouts continuations from the trace prefix before s_i; importance is KL(P_i || P_{i+1}) over bucketed answers. Model: Qwen/Qwen3-8B (revision b968826d9c46dd6066d109eabc6255188de91218), thinking on.

Parameters 

{
  "sampling": {
    "temperature": 0.9,
    "top_p": 0.95,
    "top_k": 20,
    "min_p": 0.0,
    "max_tokens": 32768,
    "seed": "not sent"
  },
  "rollout_caps": {
    "max_tokens": 16384,
    "margin": 64,
    "min_room": 256
  },
  "n_rollouts": 100,
  "n_instances": 10,
  "max_sentences": 400
}

Sources and runs

Tagged traces: connections-dev/create-interpret-task-general-tagged @ 7a7f2b90ec22e5cdb5e9afce6fddf0bb4d9d4c12. Raw traces: connections-dev/create-interpret-task-general @ 4affe1fc0e49bbf908e169c827457c7a02d42fd8. Skipped instances (too many sentences): 52. Run commands:

/scratch/mw4141/code/create_interpret/anchors/cli.py rollout --tagged-repo connections-dev/create-interpret-task-general-tagged --samples-repo connections-dev/create-interpret-task-general --subset simpleqa_verified --n-instances 10 --max-sentences 400 --min-correct-frac 0.2 --max-correct-frac 0.8 --n-rollouts 100 --max-tokens 16384 --margin 64 --min-room 256 --concurrency 64 --out-dir /scratch/mw4141/anchors_out/full_simpleqa_verified --push-to-hub --push-org connections-dev --push-repo-name create-interpret-task-general-anchors --push-visibility private --config configs/general_full.json
/scratch/mw4141/code/create_interpret/anchors/cli.py rollout --tagged-repo connections-dev/create-interpret-task-general-tagged --samples-repo connections-dev/create-interpret-task-general --subset simpleqa_verified --n-instances 10 --max-sentences 400 --min-correct-frac 0.2 --max-correct-frac 0.8 --n-rollouts 100 --max-tokens 16384 --margin 64 --min-room 256 --concurrency 64 --out-dir /scratch/mw4141/anchors_out/full_simpleqa_verified --push-to-hub --push-org connections-dev --push-repo-name create-interpret-task-general-anchors --push-visibility private --config configs/general_full.json

git sha(s): 3df492e12210930d6418b45f2ec1cd9a6f05ee42, 3df492e12210930d6418b45f2ec1cd9a6f05ee42. Full manifest: manifests/simpleqa_verified.json.

simpleqa_verified_rollouts columns

Column Description
instance_id / sample_idx instance and the index of the tagged base trace the prefix was cut from
sentence_idx the sentence the prefix stops before (the model writes its own replacement for it)
rollout_idx 0..n_rollouts-1 within the set
content the continuation only (starts right after the prefix; thinking, , answer inline)
finish_reason / success / error vLLM finish_reason; false if the request failed
max_tokens / prompt_tokens per-request generation cap and prefix length in tokens
request_usage usage block of the whole n-sample request

simpleqa_verified_positions columns

Column Description
sentence_idx / sentence / tags position, its text and function tags from the tagged repo
n_sentences / rel_pos trace length and sentence_idx / (n_sentences - 1)
prefix_chars / prompt_tokens / max_tokens prefix length (chars, tokens) and the cap used
status / reason ok, or refused (not enough room in the context window) with the reason

simpleqa_verified_importance columns

Column Description
counts bucketed answer counts of the set at this sentence ( = no extractable answer)
n_ok / n_failed / n_no_answer completions used / failed requests / no extractable answer
top_answer most common bucket
importance KL(P_i
alpha smoothing used

rat

Rollouts: 133400 rows written (133400 successful after retries, 0 still failed). finish_reason counts: {'stop': 133400}. Positions: 1334 sentences over 10 instances, 0 refused. Importance: 1324 rows, 1324 scored, alpha=0.1, mean 0.0858.

Method: Thought Anchors resampling importance. For sentence i, the set is n_rollouts continuations from the trace prefix before s_i; importance is KL(P_i || P_{i+1}) over bucketed answers. Model: Qwen/Qwen3-8B (revision b968826d9c46dd6066d109eabc6255188de91218), thinking on.

Parameters 

{
  "sampling": {
    "temperature": 0.9,
    "top_p": 0.95,
    "top_k": 20,
    "min_p": 0.0,
    "max_tokens": 32768,
    "seed": "not sent"
  },
  "rollout_caps": {
    "max_tokens": 16384,
    "margin": 64,
    "min_room": 256
  },
  "n_rollouts": 100,
  "n_instances": 10,
  "max_sentences": 400
}

Sources and runs

Tagged traces: connections-dev/create-interpret-task-general-tagged @ 7a7f2b90ec22e5cdb5e9afce6fddf0bb4d9d4c12. Raw traces: connections-dev/create-interpret-task-general @ 4affe1fc0e49bbf908e169c827457c7a02d42fd8. Skipped instances (too many sentences): 39. Run commands:

/scratch/mw4141/code/create_interpret/anchors/cli.py rollout --tagged-repo connections-dev/create-interpret-task-general-tagged --samples-repo connections-dev/create-interpret-task-general --subset rat --n-instances 10 --max-sentences 400 --min-correct-frac 0.2 --max-correct-frac 0.8 --n-rollouts 100 --max-tokens 16384 --margin 64 --min-room 256 --concurrency 64 --out-dir /scratch/mw4141/anchors_out/full_rat --push-to-hub --push-org connections-dev --push-repo-name create-interpret-task-general-anchors --push-visibility private --config configs/general_full.json

git sha(s): 3df492e12210930d6418b45f2ec1cd9a6f05ee42. Full manifest: manifests/rat.json.

rat_rollouts columns

Column Description
instance_id / sample_idx instance and the index of the tagged base trace the prefix was cut from
sentence_idx the sentence the prefix stops before (the model writes its own replacement for it)
rollout_idx 0..n_rollouts-1 within the set
content the continuation only (starts right after the prefix; thinking, , answer inline)
finish_reason / success / error vLLM finish_reason; false if the request failed
max_tokens / prompt_tokens per-request generation cap and prefix length in tokens
request_usage usage block of the whole n-sample request

rat_positions columns

Column Description
sentence_idx / sentence / tags position, its text and function tags from the tagged repo
n_sentences / rel_pos trace length and sentence_idx / (n_sentences - 1)
prefix_chars / prompt_tokens / max_tokens prefix length (chars, tokens) and the cap used
status / reason ok, or refused (not enough room in the context window) with the reason

rat_importance columns

Column Description
counts bucketed answer counts of the set at this sentence ( = no extractable answer)
n_ok / n_failed / n_no_answer completions used / failed requests / no extractable answer
top_answer most common bucket
importance KL(P_i
alpha smoothing used
Downloads last month
60
Number of rows:
138,086
Total file size:
813 MB