Dataset Viewer
Auto-converted to Parquet Duplicate
dataset_version
stringclasses
1 value
is_primary
bool
1 class
id
stringlengths
11
11
story
stringlengths
1.18k
14.1k
prompt
stringlengths
238
503
system_prompt
stringclasses
1 value
prompt_provenance_available
bool
1 class
strategy_id
stringclasses
4 values
complexified
bool
1 class
genre_path
stringclasses
27 values
tone_path
stringclasses
38 values
length_path
stringclasses
40 values
pov_path
stringclasses
27 values
structure_path
stringclasses
39 values
setting_path
stringclasses
36 values
medium_path
stringclasses
32 values
story_embedding
list
prompt_embedding
listlengths
768
768
embedding_model
stringclasses
1 value
embedding_dimensions
int16
768
768
requested_generation_model
stringclasses
1 value
generation_temperature
float32
1
1
generation_max_tokens
int32
8.19k
8.19k
generation_reasoning_effort
stringclasses
1 value
generation_model
stringclasses
1 value
finish_reason
stringclasses
1 value
generation_prompt_tokens
int32
404
454
generation_completion_tokens
int32
284
3.15k
generation_total_tokens
int32
709
3.58k
story_words
int32
213
2.36k
sentences
int16
1
183
paragraphs
int16
1
94
mean_sentence_words
float32
5.34
442
mean_paragraph_words
float32
10.1
786
mattr_500
float32
0.43
0.73
flesch_reading_ease
float32
-363.91
96.4
dialogue_marks_per_1k
float32
0
167
em_dashes_per_1k
float32
0
97
output_second_person_heuristic
bool
2 classes
document_format_heuristic
bool
2 classes
semantic_cluster
int8
0
3
nearest_story_id
stringlengths
11
11
nearest_story_cosine
float32
0.77
0.93
prompt_story_cosine
float32
0.62
0.9
source_prompt_retrieval_rank
int16
1
25
target_words_low
int16
200
1.1k
target_words_high
int16
250
1.4k
within_requested_range
bool
2 classes
target_source
stringclasses
2 values
v2
true
story-00000
The trouble started in the cobbler's square in Ardan, where Sera had spent three mornings pretending to sell ribbons while watching the palace gates. When Venn finally came out, he came out running, and the bells came out after him. "You said you'd wait for my signal," she said, once they'd put two alleys and a laundr...
Write a medium-length epic-fantasy story of roughly 800–1100 words about two former companions pursued by a royal sorcerer after one steals a relic capable of ending a long war. They shelter in libraries, courthouses, and temple courtyards across the realm, growing close again under pressure before the survivor must re...
You are an excellent short-story writer. Your stories are creative and alive — inventive where it counts, clean and assured on the page, the kind people finish and want to read again. Write only the story — no title, preamble, author note, or commentary. Follow the brief (genre, tone, POV, structure, length band). Whe...
true
speculative-and-suspense
false
genre / speculative fiction / secondary-world and epic fantasy
tone / tense_and_suspenseful / pursuit_and_escape
length / medium multi-beat arc / medium multi-beat romance or relationship
person / first_person / retrospective_protagonist
structure / circular structure / extended-span homecoming circle
setting / contemporary-urban / civic-and-public-spaces
medium / conventional narrated prose / third-person limited single-arc
[ 0.040904927998781204, 0.006832833867520094, -0.018179014325141907, 0.015921171754598618, -0.012390808202326298, -0.06146511062979698, -0.03630303964018822, 0.008762448094785213, -0.011937959119677544, -0.1381777971982956, 0.00246817315928638, 0.015759414061903954, 0.02559114620089531, 0.03...
[ 0.02205018699169159, 0.00918840616941452, -0.04529929906129837, 0.004271159414201975, -0.023318080231547356, -0.02610795386135578, 0.02325485274195671, 0.014963311143219471, -0.01153914537280798, -0.1070333868265152, -0.05130261555314064, 0.0361267514526844, 0.07181625068187714, 0.01298967...
google/gemini-embedding-2-preview
768
moonshotai/Kimi-K3
1
8,192
low
accounts/fireworks/models/kimi-k3
stop
434
1,320
1,754
1,040
72
27
14.444445
38.51852
0.5336
80.485619
26.923077
8.653846
false
false
3
story-00264
0.874214
0.756881
3
800
1,100
true
exact_prompt
v2
true
story-00001
The bus from Saskatoon let me off at the gravel turnoff, same as always, and I walked the last two miles with my suitcase bumping along behind me like a stubborn dog. I'd been away eleven years. Eleven years of apartment lobbies and takeout containers, of telling colleagues at the insurance firm that I grew up "on a fa...
For a medium-length story, follow a woman returning to her family’s prairie orchard after years in the city, where the trees still bear fruit containing brief memories of the dead. In her first-person retrospective account, nostalgia for childhood harvests and a beloved relative’s presence is complicated by the realiza...
You are an excellent short-story writer. Your stories are creative and alive — inventive where it counts, clean and assured on the page, the kind people finish and want to read again. Write only the story — no title, preamble, author note, or commentary. Follow the brief (genre, tone, POV, structure, length band). Whe...
true
speculative-and-suspense
false
genre / speculative fiction / contemporary and urban fantasy
tone / bittersweet_and_melancholic / wistful_nostalgia
length / medium extended arc / medium extended arc literary and slice-of-life
person / close_limited_third / retrospective_deep_interiority
structure / twist or reveal structure / unreliable-account reveal
setting / rural-small-town / farmland-and-prairie
medium / conventional narrated prose / first-person retrospective narration
[ 0.03162137046456337, 0.013252065517008305, -0.023661667481064796, 0.015907656401395798, -0.0009603650541976094, -0.039006758481264114, -0.03450692817568779, -0.024512195959687233, 0.025120580568909645, -0.11432518064975739, 0.021364431828260422, 0.015562856569886208, 0.031115638092160225, ...
[ 0.021402563899755478, 0.026686860248446465, -0.006199506111443043, 0.00046785513404756784, 0.014981507323682308, -0.015749163925647736, -0.010524219833314419, -0.027073470875620842, 0.021041132509708405, -0.07542628049850464, -0.010867404751479626, 0.01762321963906288, 0.008240161463618279, ...
google/gemini-embedding-2-preview
768
moonshotai/Kimi-K3
1
8,192
low
accounts/fireworks/models/kimi-k3
stop
427
1,693
2,120
1,344
85
19
15.811765
70.736839
0.53185
77.860168
8.928572
6.696429
false
false
1
story-00752
0.874589
0.789297
1
1,100
1,400
true
taxonomy_length_description
v2
true
story-00002
I was fifty-one when they came to my door, which is too old for marsh work, and I told them so. Tellan didn't argue. He just set the shrine token on my table—a clam shell with a hole drilled through it, the kind they hang in the drowned villages—and said the bell had been seen. Wreckers had it, down in the Greywater, s...
In a low-magic salt-marsh kingdom, a retired pathfinder joins a small expedition to recover the bell taken from a drowned village shrine, whose sound can briefly guide the dead home. Tell this as a 600–750-word speculative scene in first-person retrospective narration, centering the cost of carrying one lost companion’...
You are an excellent short-story writer. Your stories are creative and alive — inventive where it counts, clean and assured on the page, the kind people finish and want to read again. Write only the story — no title, preamble, author note, or commentary. Follow the brief (genre, tone, POV, structure, length band). Whe...
true
period-folk-and-adventure
false
genre / adventure, humor, and traditional modes / adventure and quest
tone / bittersweet_and_melancholic / elegiac_grief
length / short scene with small arc / speculative scene
person / third_omniscient / neutral_panoramic
structure / circular structure / the cycle repeats with a new figure
setting / secondary-invented-world / low-magic-grounded-realm
medium / conventional narrated prose / first-person retrospective narration
[ 0.04973715916275978, 0.032473187893629074, -0.020168505609035492, -0.006288430187851191, -0.007773140911012888, -0.05289063975214958, -0.03420126438140869, 0.01994556188583374, -0.011507752351462841, -0.11562293767929077, 0.02202640101313591, 0.019698316231369972, 0.00951480958610773, 0.01...
[ 0.03577970713376999, 0.012155521661043167, -0.035657621920108795, -0.018206970766186714, -0.015406569465994835, -0.054055262356996536, 0.016061531379818916, -0.02583669126033783, 0.015286002308130264, -0.10632213950157166, -0.000002391886255281861, 0.009340083226561546, 0.010065203532576561,...
google/gemini-embedding-2-preview
768
moonshotai/Kimi-K3
1
8,192
low
accounts/fireworks/models/kimi-k3
stop
429
976
1,405
804
40
12
20.1
67
0.4999
81.525291
0
6.218905
false
false
2
story-00040
0.904726
0.777642
1
600
750
false
exact_prompt
v2
true
story-00003
The bell above the door hadn't rung in two hours when Rosa pulled out the unclaimed bin. It was a Tuesday, the kind of gray afternoon that made the shop feel like a waiting room for a train that wasn't coming. She'd already re-alphabetized the thread drawer and restitched the hem of her own cardigan twice. The bin was...
During a slow afternoon at a neighborhood alterations shop, a seamstress reluctantly helps a former customer identify a coat left unclaimed for years, only to discover halfway through the conversation that the garment connects to someone she thought had simply moved away. Write a 600–750-word literary slice-of-life sce...
You are an excellent short-story writer. Your stories are creative and alive — inventive where it counts, clean and assured on the page, the kind people finish and want to read again. Write only the story — no title, preamble, author note, or commentary. Follow the brief (genre, tone, POV, structure, length band). Whe...
true
everyday-default
false
genre / literary and everyday life / slice-of-life vignette
tone / bittersweet_and_melancholic / wistful_nostalgia
length / short scene with small arc / literary slice-of-life scene
person / close_limited_third / retrospective_measured_distance
structure / twist or reveal structure / midpoint pivot reveal
setting / contemporary-urban / work-and-commerce
medium / conventional narrated prose / third-person limited single-arc
[ 0.041284386068582535, -0.006062425673007965, 0.050315242260694504, -0.0248090922832489, -0.028972681611776352, 0.006850881036370993, -0.019485345110297203, 0.02525230124592781, 0.03229222819209099, -0.1188756600022316, 0.005830568727105856, -0.001588383805938065, 0.022091343998908997, 0.02...
[ 0.05489976331591606, -0.027442334219813347, 0.033843033015728, -0.0525667704641819, -0.04042818769812584, 0.017301583662629128, -0.014512798748910427, 0.018555693328380585, 0.020602358505129814, -0.08444935828447342, 0.009192650206387043, -0.0035840559285134077, 0.0290694460272789, 0.03829...
google/gemini-embedding-2-preview
768
moonshotai/Kimi-K3
1
8,192
low
accounts/fireworks/models/kimi-k3
stop
429
1,060
1,489
804
93
24
8.645162
33.5
0.5078
83.576576
57.213932
11.19403
false
false
1
story-00568
0.859157
0.76289
1
600
750
false
exact_prompt
v2
true
story-00004
"The Saturday line was out the door by nine, and somewhere between the man who wanted his sourdough (...TRUNCATED)
"During a crowded Saturday shift at a neighborhood bakery, a new cashier argues with her older cowor(...TRUNCATED)
"You are an excellent short-story writer. Your stories are creative and alive — inventive where it(...TRUNCATED)
true
everyday-default
false
genre / literary and everyday life / slice-of-life vignette
tone / warm_and_hopeful / quietly_uplifting
length / short scene with small arc / comedy whimsy scene
person / first_person / retrospective_protagonist
structure / in medias res opening / heated confrontation opening with relational stakes
setting / contemporary-urban / work-and-commerce
medium / conventional narrated prose / third-person limited single-arc
[0.050703030079603195,0.021821340546011925,0.005442507565021515,-0.011731402948498726,0.004079016391(...TRUNCATED)
[0.04115857183933258,-0.0025327797047793865,-0.02097499556839466,-0.020821379497647285,0.00772057846(...TRUNCATED)
google/gemini-embedding-2-preview
768
moonshotai/Kimi-K3
1
8,192
low
accounts/fireworks/models/kimi-k3
stop
422
1,103
1,525
794
90
36
8.822222
22.055555
0.52455
79.717476
98.236778
5.037783
false
false
1
story-01377
0.880008
0.70684
2
600
750
false
exact_prompt
v2
true
story-00005
"The winter I was eleven, packages came to our street the way news came—slowly, and then all at on(...TRUNCATED)
"Write a 600–750-word fable scene, told by an adult looking back, in which I find a stray dog guar(...TRUNCATED)
"You are an excellent short-story writer. Your stories are creative and alive — inventive where it(...TRUNCATED)
true
everyday-default
false
genre / literary and everyday life / literary character drama
tone / warm_and_hopeful / warm_and_tender
length / short scene with small arc / fable or folktale scene
person / first_person / retrospective_protagonist
structure / linear rising arc with resolution / single-scene compressed arc
setting / contemporary-urban / residential-neighborhoods
medium / conventional narrated prose / first-person retrospective narration
[-0.0011200401932001114,0.015843801200389862,-0.034727297723293304,-0.017774362117052078,-0.00231104(...TRUNCATED)
[0.0004397184820845723,-0.031905051320791245,-0.030192781239748,-0.06229653209447861,0.0069521963596(...TRUNCATED)
google/gemini-embedding-2-preview
768
moonshotai/Kimi-K3
1
8,192
low
accounts/fireworks/models/kimi-k3
stop
418
945
1,363
763
52
13
14.673077
58.692307
0.4785
81.285469
10.484928
7.863696
false
false
3
story-01586
0.817759
0.819449
1
600
750
false
exact_prompt
v2
true
story-00006
"Mara found the note taped to her kettle at six in the morning, which meant Yusuf had been inside he(...TRUNCATED)
"Write a 750–900-word thriller about two neighbors attempting to recover a stolen evidence bag som(...TRUNCATED)
"You are an excellent short-story writer. Your stories are creative and alive — inventive where it(...TRUNCATED)
true
period-folk-and-adventure
false
genre / adventure, humor, and traditional modes / adventure and quest
tone / tense_and_suspenseful / paranoid_suspicion
length / short compressed journey / mystery thriller journey
person / close_limited_third / retrospective_deep_interiority
structure / linear rising arc with resolution / slow-build arc with quiet climactic realization
setting / contemporary-urban / residential-neighborhoods
medium / conventional narrated prose / third-person omniscient or roving narration
[0.05359946936368942,0.005716658663004637,0.02353353053331375,0.0055537414737045765,-0.0068520247004(...TRUNCATED)
[0.030087856575846672,0.0227067731320858,0.0016851966502144933,-0.025032889097929,-0.015505118295550(...TRUNCATED)
google/gemini-embedding-2-preview
768
moonshotai/Kimi-K3
1
8,192
low
accounts/fireworks/models/kimi-k3
stop
442
1,097
1,539
860
79
31
10.886076
27.741936
0.53075
80.985397
44.186047
2.325581
false
false
3
story-01522
0.824322
0.797634
1
750
900
true
exact_prompt
v2
true
story-00007
"The trouble began with the flowers, which Albert had wrapped in damp newspaper and stowed under the(...TRUNCATED)
"In a 1968 seaside town, a bus conductor tries to give a proper farewell to his last regular passeng(...TRUNCATED)
"You are an excellent short-story writer. Your stories are creative and alive — inventive where it(...TRUNCATED)
true
quiet-and-comic-flash
false
genre / literary and everyday life / slice-of-life vignette
tone / bittersweet_and_melancholic / bittersweet_parting
length / short scene with small arc / comedy whimsy scene
person / first_person / retrospective_protagonist
structure / circular structure / return to the opening image
setting / historical-past / recent-past-twentieth-century
medium / conventional narrated prose / quiet present-tense third person
[0.002488184254616499,0.0011011002352461219,0.004833347164094448,-0.01207839883863926,-0.01796315796(...TRUNCATED)
[0.0027123435866087675,0.001011283602565527,-0.008099716156721115,-0.02336236834526062,-0.0254898313(...TRUNCATED)
google/gemini-embedding-2-preview
768
moonshotai/Kimi-K3
1
8,192
low
accounts/fireworks/models/kimi-k3
stop
429
1,059
1,488
784
50
23
15.68
34.086956
0.54815
67.364952
35.714287
14.030612
false
false
1
story-00768
0.864546
0.742787
1
600
750
false
exact_prompt
v2
true
story-00008
"The shuttle to Dock Nine was supposed to take eleven minutes. Mother Reyes had ridden it four thous(...TRUNCATED)
"On a far-future generation ship, an elderly route conductor boards a shuttle whose artificial stree(...TRUNCATED)
"You are an excellent short-story writer. Your stories are creative and alive — inventive where it(...TRUNCATED)
true
speculative-and-suspense
false
genre / speculative fiction / far-future and spacefaring science fiction
tone / bittersweet_and_melancholic / wistful_nostalgia
length / short scene with small arc / comedy whimsy scene
person / close_limited_third / retrospective_measured_distance
structure / linear rising arc with resolution / single-scene compressed arc
setting / contemporary-urban / transit-and-streets
medium / conventional narrated prose / third-person limited single-arc
[0.0025915848091244698,0.014608322642743587,-0.016292350366711617,-0.006903466768562794,-0.029210578(...TRUNCATED)
[0.02549918182194233,0.029402997344732285,-0.0047568087466061115,-0.03475034609436989,-0.00889538042(...TRUNCATED)
google/gemini-embedding-2-preview
768
moonshotai/Kimi-K3
1
8,192
low
accounts/fireworks/models/kimi-k3
stop
431
969
1,400
740
68
25
10.882353
29.6
0.5599
73.119415
83.783783
5.405406
false
false
2
story-00594
0.852515
0.700755
1
600
750
true
exact_prompt
v2
true
story-00009
"The gate was locked, which was the first wrong thing.\n\nI keep the Bramblewick pass gate myself. I(...TRUNCATED)
"Write an 800–1100-word fantasy story in which I, a roadwarden stationed at a remote forest hamlet(...TRUNCATED)
"You are an excellent short-story writer. Your stories are creative and alive — inventive where it(...TRUNCATED)
true
period-folk-and-adventure
false
genre / speculative fiction / secondary-world and epic fantasy
tone / tense_and_suspenseful / edge_of_revelation
length / medium multi-beat arc / medium multi-beat literary
person / close_limited_third / retrospective_deep_interiority
structure / linear rising arc with resolution / single-scene compressed arc
setting / rural-small-town / mountain-and-forest-hamlet
medium / conventional narrated prose / first-person retrospective narration
[0.042353808879852295,-0.01051768846809864,0.028175026178359985,-0.021151607856154442,0.008956984616(...TRUNCATED)
[0.046962834894657135,-0.009244394488632679,0.0059668985195457935,-0.017734043300151825,-0.020215027(...TRUNCATED)
google/gemini-embedding-2-preview
768
moonshotai/Kimi-K3
1
8,192
low
accounts/fireworks/models/kimi-k3
stop
424
1,488
1,912
1,145
80
30
14.3125
38.166668
0.4822
84.285805
40.174671
3.49345
false
false
3
story-01032
0.859215
0.811596
1
800
1,100
false
exact_prompt
End of preview. Expand in Data Studio

Kimi K3 Story Corpus with Gemini Embeddings

Generated with Simula

V1 vs. V2: Use V2 for new work. V1 is the original generation built with the legacy Simula prompt taxonomy, where narration/POV and delivery medium were partly combined and second-person or document-shaped stories appeared too often. V2 is a fresh regeneration from revised Simula prompts: grammatical person/focalization and delivery medium are separated, complexification is disabled, the strategy set is simplified, and prompts are substantially shorter. The splits contain different stories under the same stable IDs; V1 is retained as a reproducible before/after baseline.

A versioned Mercity Research Labs corpus of synthetic English short stories with whole-document Gemini embeddings. V2 is the primary split. V1 remains available as a separate, schema-compatible split for reproducibility and direct before/after analysis.

Both splits contain 2,000 complete stories generated with Kimi K3 and indexed by the same stable IDs, story-00000 through story-01999. They are different story generations, not duplicate copies.

Versions and recommended split

Split Status Rows Story text Exact prompts Story vectors Prompt vectors
v2 Primary / recommended 2,000 Complete Complete 2,000 2,000
v1 Archived comparison baseline 2,000 Complete Complete 2,000 2,000

Use v2 for new story-corpus work. Both versions now include exact prompt provenance and paired prompt/story embeddings; use v1 as the archived comparison baseline.

from datasets import load_dataset

v2 = load_dataset(
    "Mercity/kimi-k3-story-corpus-embeddings",
    split="v2",
)

v1 = load_dataset(
    "Mercity/kimi-k3-story-corpus-embeddings",
    split="v1",
)

print(v2[0]["story"])
print(len(v2[0]["story_embedding"]))  # 768

What changed in V2

The regenerated V2 corpus is shorter, more conversational, less dominated by second-person narration and documentary formats, and less semantically crowded than V1.

Measure V1 V2
Median story length 984 words 776 words
Median MATTR-500 0.528 0.538
Median Flesch Reading Ease 73.9 79.4
Median nearest-story cosine 0.873 0.859
Story pairs at cosine ≥ 0.90 203 55

V2's most important remaining limitation is cross-record lexical identity reuse. Several character names and scene templates recur at conspicuous rates. See analysis/v2_deep_quality_semantic_audit.pdf for the full quantitative and human-reading audit.

Provenance and privacy boundary

  • story always contains the complete final narrative returned in message.content.
  • The response-only V2 output JSONL did not itself contain prompts. The authoritative Kimi batch input and Simula lineage were recovered from the original run and joined to all 2,000 outputs by exact custom_id.
  • Every recovered V2 batch prompt exactly matches the selected Simula meta-prompt for its recorded attempt_index; strategy and taxonomy lineage were also checked for exact equality.
  • Raw V2 responses included a private reasoning_content field. That field is intentionally excluded from this public dataset. It must not be used as story text or training text.
  • V2 numeric targets come from exact prompts where numeric wording is present; 63 qualitative-only prompts use the authoritative sampled length-taxonomy description instead.
  • Both splits include exact final prompts, their system prompts, Simula lineage, and paired whole-document prompt/story embeddings.

Shared schema

Both splits use the same 49-column schema. A nullable value is meaningful: it indicates that a metric does not apply to a row or that a taxonomy dimension did not exist in that version. It does not indicate an upload failure.

Version, identity, and story text

Column Storage type Definition and nullability
dataset_version string Explicit split version: v1 or v2. Complete in both splits.
is_primary boolean true for every V2 row and false for every V1 row. This is a release preference, not a quality label for an individual story.
id string Stable row ID from story-00000 through story-01999. IDs align corresponding V1 and V2 generation slots; matching IDs do not imply matching story content.
story string Complete, unabridged final narrative. Only message.content is stored; model reasoning is excluded. Complete in both splits.
prompt string Exact final user prompt sent to Kimi. Complete in both splits. V2 values come from the recovered authoritative batch input, not from story-text inference.
system_prompt string Exact system-level writing instruction sent alongside prompt. Complete in both splits; each split uses one shared system prompt.
prompt_provenance_available boolean true when authoritative prompt/system text and lineage were exactly joined. It is true for every published V1 and V2 row.

Prompt-construction provenance

These columns preserve the Simula prompt-construction lineage. They are source fields, not model-predicted labels. V2 deliberately separates grammatical person/focalization (pov_path) from delivery form (medium_path); V1's older taxonomy combined these ideas inside pov_path and therefore has no exact medium_path.

Column Storage type Definition
strategy_id string Exact Simula generation strategy ID. V1 uses five legacy strategies; V2 uses everyday-default, speculative-and-suspense, quiet-and-comic-flash, or period-folk-and-adventure.
complexified boolean Whether the selected brief received Simula's extra constraint-interaction pass. V2 intentionally disabled complexification, so all V2 values are false.
genre_path string Sampled genre taxonomy path, root to selected node, separated by /. Complete in both splits.
tone_path string Sampled tone taxonomy path in the same root-to-node format. Complete in both splits.
length_path string Sampled categorical target-length taxonomy path. Numeric targets are stored separately. Complete in both splits.
pov_path string Requested narration taxonomy path. In V1 this is the legacy combined POV taxonomy; in V2 it is the redesigned dedicated person/focalization path. Complete in both splits, but compare the versions with this taxonomy change in mind.
structure_path string Sampled narrative plot/temporal-structure path. Complete in both splits.
setting_path string Sampled setting taxonomy path. Complete in both splits.
medium_path nullable string V2's dedicated delivery-medium path: conventional prose, epistolary, documentary/archive, mixed frame, or another experimental form. Complete in V2 and null in V1 because the V1 taxonomy had no separate medium factor.

Embeddings

All embeddings were created from the complete source string in one request using google/gemini-embedding-2-preview through OpenRouter. The corpus pipeline did not chunk, pool, summarize, or truncate the stored text before embedding.

Column Storage type Definition and nullability
story_embedding fixed-size list of 768 float32 values Whole-document embedding of story. Complete in both splits.
prompt_embedding list of 768 float32 values Whole-document embedding of prompt. Complete in both splits. The Parquet storage is a nullable list for cross-tool compatibility, but validation confirms exactly 768 finite values in every published row.
embedding_model string Exact model identifier: google/gemini-embedding-2-preview. Complete in both splits.
embedding_dimensions int16 Stored vector dimensionality. Always 768.

For cosine search, L2-normalize vectors before taking a dot product. Embedding proximity measures semantic relatedness, not literary merit, textual duplication, or factual correctness.

Generation metadata

Request fields come from the authoritative batch inputs; response and usage fields come from the generation API output.

Column Storage type Definition
requested_generation_model string Model identifier placed in the batch request: moonshotai/Kimi-K3. Complete in both splits.
generation_temperature float32 Sampling temperature from the batch request. It is 1.0 for every published row.
generation_max_tokens int32 Maximum completion-token budget from the request: 4,096 in V1 and 8,192 in V2. It is a ceiling, not actual output usage.
generation_reasoning_effort nullable string Requested reasoning-effort setting. It is low for every V2 request and null in V1, whose batch format did not include this option. No reasoning trace is published.
generation_model string Model identifier returned by the generation provider: accounts/fireworks/models/kimi-k3.
finish_reason string Provider finish reason for the selected response choice. All published rows completed with stop.
generation_prompt_tokens int32 Provider-reported input-token count for the generation request.
generation_completion_tokens int32 Provider-reported output-token count. It may differ from story_words because model tokens and regex words are different units.
generation_total_tokens int32 Provider-reported total tokens for the request.

Quantitative language measurements

Word measurements use the case-insensitive pattern [A-Za-z]+(?:['’][A-Za-z]+)?: alphabetic tokens with an optional internal apostrophe. Numbers and punctuation are not words. Sentence, paragraph, syllable, dialogue, and document-format fields are deterministic heuristics rather than human annotations.

Column Storage type Definition
story_words int32 Number of regex word tokens in story. Observed range: V1 228–2,732; V2 213–2,362.
sentences int16 Non-empty sentence segments detected from punctuation-aware rule-based splitting. Abbreviations and experimental formatting can affect the count.
paragraphs int16 Non-empty blocks separated by one or more blank lines. A single line break does not create a new paragraph.
mean_sentence_words float32 Arithmetic mean regex-word count across detected non-empty sentence segments. Unit: words per sentence.
mean_paragraph_words float32 Arithmetic mean regex-word count across non-empty blank-line-delimited paragraphs. Unit: words per paragraph.
mattr_500 float32 Moving-Average Type–Token Ratio using a 500-word window. Each sampled window contributes unique lowercased tokens / window tokens; up to 40 evenly spaced windows are averaged. Stories at or below 500 tokens use one whole-story ratio. The 500 is the window length, not a score or sample count.
flesch_reading_ease float32 Approximate English Flesch Reading Ease, 206.835 - 1.015 × words/sentences - 84.6 × syllables/words, using heuristic sentence and syllable counts. Use comparatively; unusual formatting can produce extreme values.
dialogue_marks_per_1k float32 Count of straight and curly quotation marks divided by story_words and multiplied by 1,000. This approximates dialogue density but can also count quotations or stylized punctuation.
em_dashes_per_1k float32 Unicode em-dash count divided by story_words and multiplied by 1,000.

Narrative-form heuristics

Column Storage type Definition
output_second_person_heuristic boolean Conservative output-text signal. Quoted spans are removed, then the row is true when you, your, yours, yourself, or yourselves occur at least 35 times per 1,000 remaining word tokens. This detects strong second-person narration, not every incidental use of “you.”
document_format_heuristic boolean true when line openings match document-like markers such as all-cap headers, exhibits, letters, dated entries, or numbered-day labels. This is a formatting signal, not a genre annotation.

Story-semantic fields

Cosines are calculated after L2-normalizing story vectors. Cluster IDs and nearest neighbors are computed independently within each version; compare values within a split unless explicitly conducting a version study.

Column Storage type Definition
semantic_cluster int8 Exploratory spherical-k-means cluster ID, 0–3, remapped by descending cluster size within the split. IDs are version-specific and are not ground-truth genre labels.
nearest_story_id string id of the most cosine-similar other story in the same split. Self-matches are excluded.
nearest_story_cosine float32 Cosine similarity to nearest_story_id. Higher means greater embedding proximity, not duplication or higher quality.

Prompt alignment and length targets

Column Storage type Definition and nullability
prompt_story_cosine float32 Cosine between the row's whole-story and exact whole-prompt embedding. Complete in both splits. V2 observed median: 0.787.
source_prompt_retrieval_rank int16 One-based rank of the true source prompt when the story is searched against all 2,000 prompt vectors in the same split. Rank 1 means the source prompt is closest. Complete in both splits; V2 top-1 retrieval is 94.9% and top-10 is 99.8%.
target_words_low nullable int16 Lower numeric target. Parsed from the exact prompt when numeric wording is present; otherwise V2 falls back to the selected authoritative length-taxonomy description. V2 is complete; V1 has 305 null rows with no supported numeric range.
target_words_high nullable int16 Upper numeric target from the same source as target_words_low. For an approximate single target, low and high are equal.
within_requested_range nullable boolean Inclusive range check. It is null for 11 V2 rows that state an approximate single target rather than a range, and for 305 V1 rows without a parsed range.
target_source nullable string Provenance of the numeric target: exact_prompt or, for 63 qualitative-only V2 prompts, taxonomy_length_description. Null only on the 305 V1 rows without a numeric target.

Nullability summary

Field group V1 V2
Exact prompt, system prompt, strategy, shared taxonomy paths Complete Complete
Story and prompt embeddings; prompt alignment Complete Complete
V2-only medium_path 2,000 null by design Complete
Numeric target fields 305 null Complete
Literal within_requested_range 305 null 11 null for approximate single targets
Story text and story embeddings Complete Complete

Quantitative overview

  • 4,000 rows total across two splits; 2,000 unique IDs and 2,000 complete story vectors per split.
  • No empty stories or exact duplicate story text within either version.
  • V2 contains 1,390,446 regex word tokens; its median story length is 776 words.
  • V2's median MATTR-500 is 0.538 and median Flesch Reading Ease is 79.4.
  • V2 median nearest-story cosine is 0.859, down from 0.873 in V1.
  • V2 has 55 story pairs at cosine ≥ 0.90, compared with 203 in V1.
  • V2 median exact prompt–story cosine is 0.787; source-prompt retrieval is 94.9% at rank 1 and 99.8% within the top 10.
  • V2 contains 1,926 exact numeric prompt ranges; literal compliance is 39.4%, with a median output/midpoint ratio of 1.135.
  • V2 story-plus-prompt embedding used 1,992,956 input tokens. At the $0.20-per-million rate used during the audit, the estimated combined charge is $0.399; this is an estimate, not a billing receipt.

Analysis files

  • analysis/v2_deep_quality_semantic_audit.pdf — primary 21-page V2 prompt-fidelity, quality, language, n-gram, semantic, cost, and human-reading audit.
  • analysis/v1_story_prompt_embedding_audit.pdf — archived V1 prompt/story alignment and embedding report.

Intended uses

  • Whole-story semantic search, clustering, and recommendation.
  • Creative-writing corpus analysis and synthetic-data quality evaluation.
  • Direct V1/V2 studies of narrative form, length, readability, and semantic crowding.
  • Prompt–story alignment and source-prompt retrieval research in either version.
  • Train/evaluation split design, with high-similarity neighborhoods grouped to reduce leakage.

Limitations

  • All stories are synthetic and inherit model and pipeline biases.
  • V2 has material character-name reuse and recurring scene templates; consult the V2 audit before training.
  • V1 and V2 use different prompt-taxonomy designs. In particular, V2 separates person and medium; do not treat taxonomy paths as identical label spaces across versions.
  • Semantic similarity does not establish plagiarism, factual consistency, safety, or literary quality.
  • The four clusters are exploratory and version-specific.
  • Sentence boundaries, syllables, document formatting, second-person form, and word targets use deterministic heuristics.
  • Whole-story embeddings are unsuitable for passage-level retrieval without additional chunk embeddings.
  • google/gemini-embedding-2-preview is a preview identifier and may later change or be retired.

Recommended release handling

V2 is the primary corpus, but Mercity recommends a targeted editorial pass before model training: repair high-frequency character names, reject incomplete endings, remove explicit thesis restatements, and group high-similarity neighbors on the same side of evaluation splits. Raw model reasoning should remain private.

License

No open-data license has been declared. The card therefore uses license: other. Contact Mercity Research Labs before redistribution, commercial use, or publication of derivatives. Mercity may replace this notice with a specific license in a later revision.

Citation

@dataset{mercity_kimi_k3_story_corpus_2026,
  author       = {Mercity Research Labs},
  title        = {Kimi K3 Story Corpus with Gemini Embeddings},
  year         = {2026},
  publisher    = {Hugging Face},
  organization = {Mercity},
  note         = {V2 primary split; V1 archived comparison split}
}

Contact

Mercity Research Labs · Simula

Downloads last month
339