CoolFace
Datasetpublic

gszauer/Clue250K

Clue 250K Synthetic Corpus Web Demo Source on Github Model Page This dataset contains synthetic short murder mysteries for training and testing the tiny Clue 250K language model. The examples use a fixed set of names, locations, weapons, and wound descriptions. Each mystery asks the model to infer the murderer, or answer Unknown when the clues do not identify exactly one person. The companion code and trained demo model live in the GitHub repository for Clue 250K. The dataset… See the full description on the dataset page: https://huggingface.co/datasets/gszauer/Clue250K.

sourceHugging Faceotherupdated 4mo agoView on Hugging Face
0likes8downloads
Dataset Card

Clue 250K Synthetic Corpus

This dataset contains synthetic short murder mysteries for training and testing the tiny Clue 250K language model. The examples use a fixed set of names, locations, weapons, and wound descriptions. Each mystery asks the model to infer the murderer, or answer Unknown when the clues do not identify exactly one person.

The companion code and trained demo model live in the GitHub repository for Clue 250K. The dataset is intended for educational language-model experiments, not as a benchmark for general reasoning.

Dataset Structure

Each .txt file contains many examples separated by a blank line. There is no JSON, CSV, or metadata sidecar required to use the corpus.

Example:

text
Josh is in the library
Basil is in the kitchen
A body was found in the library
Maria is in the living room
Therefore the murderer is: Josh FIN.

FIN. is literal text in the corpus. It is not a special token.

Size

The current upload contains:

  • —Total examples: 3,164,051
  • —Total text size: 524,288,508 bytes (500.0 MiB)
  • —Total text shards: 318
  • —Location examples: 1,196,152 in 120 files
  • —Weapon examples: 1,091,322 in 110 files
  • —Unsolvable examples: 876,577 in 88 files
  • —Three-suspect examples: about 25%
  • —Four-suspect examples: about 75%

Example Types

Location Mysteries

The body location identifies the murderer.

text
Josh is in the library
Basil is in the kitchen
Maria is in the living room
A body was found in the library
Therefore the murderer is: Josh FIN.

If no person is in the body location, or more than one person is in the body location, the correct answer is Unknown.

Weapon Mysteries

The wound identifies the matching weapon, and the weapon identifies the murderer.

text
Josh has a pipe
Basil has a knife
Maria has a rope
A body was found in the library with a knife wound
Therefore the murderer is: Basil FIN.

Some weapon examples include both room and weapon information:

text
Josh is in the library with a pipe

The compact Name has a weapon form is useful for very small models with short context windows.

If no weapon matches the wound, or more than one person has the matching weapon, the correct answer is Unknown.

Unsolvable Mysteries

Unsolvable examples are generated in both location and weapon styles. They include either no valid match or an ambiguous match.

text
Josh is in the library
Basil is in the library
A body was found in the library
Maria is in the kitchen
Therefore the murderer is: Unknown FIN.

Endings

Most examples use:

text
Therefore the murderer is: {answer} FIN.

The corpus also includes alternate ending styles so a model does not only learn one final-line prefix:

text
The murderer is {answer} FIN.
The murderer was {answer} FIN.
The Murderer is {answer} FIN.
So the murderer is {answer} FIN.
So the murderer was {answer} FIN.

Approximate ending mix:

  • —Therefore the murderer is: - 70%
  • —The murderer is - 10%
  • —The murderer was - 8%
  • —The Murderer is - 5%
  • —So the murderer is - 4%
  • —So the murderer was - 3%

Intended Use

This corpus was built to overtrain a very small decoder-only transformer:

  • —Vocabulary size: 512
  • —Context length: 96
  • —Blocks: 3
  • —Feature dimension: 72
  • —Parameters in the trained reference model: 230,904

The reference training run used a byte-BPE tokenizer and produced:

  • —Byte-BPE tokens per epoch: 240,244,946
  • —Epochs: 4
  • —Total token presentations: 960,979,784
  • —Tokens per parameter per epoch: 1,040.45
  • —Total token presentations per parameter: 4,161.82

Valid Names

text
Alex, Avery, Basil, Blair, Blake, Casey, Dana, Devin, Drew, Eden, Ellis,
Emery, Erin, Felix, Finn, Gail, Gray, Harper, Hazel, Ira, Jamie, Jordan,
Josh, Jules, Kai, Kendall, Lane, Laurel, Lee, Logan, Mara, Maria, Micah,
Morgan, Nico, Noel, Nora, Owen, Paige, Parker, Peyton, Quinn, Reese, Remy,
Riley, Robin, Rowan, Sage, Sam, Sasha, Sidney, Sky, Sloan, Talia, Taylor,
Theo, Vale, Wren, Zane, Mina, Ivy, Lena, Toby, June

Valid Locations

text
library, kitchen, living room, study, hallway, cellar, attic, garden, garage,
pantry, office, lounge, chapel, conservatory, ballroom, dining room, parlor,
nursery, balcony, courtyard, observatory, gallery, theater, basement

Valid Weapons

text
knife, pipe, rope, wrench, candle, poison, revolver, hammer, axe, poker,
dagger, chain, bat, statue, scissors, shovel, vial, club, cable, needle

Valid Wounds

text
knife     -> knife wound
pipe      -> pipe bruise
rope      -> rope mark
wrench    -> wrench bruise
candle    -> candle wax
poison    -> poison trace
revolver  -> bullet wound
hammer    -> hammer dent
axe       -> axe wound
poker     -> poker burn
dagger    -> dagger wound
chain     -> chain mark
bat       -> bat bruise
statue    -> statue blow
scissors  -> scissor cut
shovel    -> shovel mark
vial      -> glass cut
club      -> club bruise
cable     -> cable mark
needle    -> needle mark

Caveats

  • —The data is highly templated.
  • —Names, locations, weapons, and wounds are intentionally constrained.
  • —The corpus intentionally contains many repeated structural patterns because it targets very small language models.