geekyrakshit/random-relationships

Random Relationships

A reproducible synthetic dataset of 500 country–capital–person triples for controlled knowledge-injection, model-editing, and relational-reasoning experiments. All entities are generated pseudonames, and all associations are artificial. The dataset contains no factual claims about real countries, cities, or people.

This release was generated for the weight-space-knowledge-injection project, which investigates whether weight-space programming can add knowledge to a model without degrading its existing capabilities. The dataset itself provides neither model-editing results nor evidence that this hypothesis holds.

Structure

Each row defines two artificial facts: the country's capital is capital, and person lives in country.

Field	Type	Meaning
country	string	Unique pseudoname for a country.
capital	string	Unique pseudoname assigned as that country's capital.
person	string	Unique pseudoname for a person assigned to live in that country.

The first record is:

{"country": "Pimame", "capital": "Godradis", "person": "Tekisor"}

It can be verbalized as “Pimame's capital is Godradis” and “Tekisor lives in Pimame.” Question wording, answers, prompts, and tokenization are left to the experimenter.

There are 500 unique countries, 500 unique capitals, and 500 unique people. All 1,500 names are globally unique across the three columns. Every country has exactly one capital and one person; every capital and person occurs once. Names are capitalized, contain only ASCII letters, and have 5–10 characters. No missing values occur in this release.

The single eval split contains all 500 original rows in their original order. The split is designated for evaluation. The generation command did not create separate training, validation, or test partitions.

Load the dataset

from datasets import load_dataset

dataset = load_dataset("geekyrakshit/random-relationships", split="eval")
print(dataset[0])

Data is published in Parquet format for standard Hub loading and browsing. provenance.json records generation parameters, the original saved-dataset fingerprint, file checksums, and the publication environment. generation.py is an exact copy of the generator source used to verify this release.

Generation and reproducibility

The original command was:

weight-space-ki generate-dataset \
  --output data/relationships \
  --n 500 \
  --seed 42 \
  --relationship-seed 42

The output directory must be new. This release explicitly uses relationship_seed=42; the generator's default is 43.

In Python, the equivalent project API call is:

from weight_space_ki import generate_dataset

dataset = generate_dataset(n=500, seed=42, relationship_seed=42)

The name seed initializes a local random generator, which supplies independent seeds for three name pools. Names combine two or three consonant–vowel syllables with an optional consonant ending. Candidates outside the length bounds or already present in any pool are rejected. A separate local generator, initialized with the relationship seed, shuffles the capital pool and then the person pool. These generators do not modify Python's global random state.

Changing only relationship_seed preserves all name pools and country order while changing assignments. Keep both seeds and the generator implementation fixed when reproducing this release.

To use the bundled generator snapshot independently of the project package, download generation.py and run:

from generation import generate_dataset

dataset = generate_dataset(n=500, seed=42, relationship_seed=42)

It requires Python 3.11+ and the datasets package. The locally saved Arrow data was checked against freshly generated records using the stated parameters, and the published Parquet records were checked for exact row-for-row equality.

Intended use and evaluation

Use the triples as controlled associations for testing factual recall, inverse relations, multi-step reasoning, and knowledge injection or editing. For example, combining the two facts in a row supports asking for the capital of the country where the named person lives.

Measure the unedited model on these associations before intervention. Construct and disclose experimental splits, paraphrases, and evaluation protocols. When measuring transfer across entities, split by row or entity before producing question variants to avoid leaking the same relationship across partitions. Measure retention of existing capabilities using separate evaluation data; this dataset does not contain a capability-retention benchmark.

Limitations

• Random associations do not establish that every name or inferred answer is absent from pretraining. Generated pseudonames can coincide with real names or words, and a model may answer correctly by chance.
• Pseudoname token lengths depend on the target tokenizer. Measure and balance tokenization across experimental conditions when relevant.
• The syllable inventory is hand-designed and does not represent the linguistic or demographic diversity of real names.
• One-to-one mappings and only 500 rows provide a deliberately simple setting. Results need not generalize to ambiguous, many-to-many, or real-world knowledge.
• No human annotation, web scraping, or external factual source is involved in generating the triples. Incidental name matches do not describe real people.

License

No dataset license was specified in the source dataset metadata or the project files inspected for this release. No license is assigned by this card. Contact the repository owner to clarify reuse terms.