Datasets:

dimhold
/

fixture-entropy

license: mit
language:
  - en
pretty_name: fixture-entropy
size_categories:
  - 1K<n<10K
task_categories:
  - text-generation
tags:
  - llm-evaluation
  - test-data
  - diversity
  - claude
  - synthetic
configs:
  - config_name: calls
    data_files:
      - split: train
        path: out/calls.jsonl

fixture-entropy

Author: Dmitriy Semenkevich (dimhold)

This dataset measures how often 2 independent test fixture sets generated by a Claude model share a value, for 7 value types across 7 Anthropic models, against Faker and true random baselines.

Headline, from the repository README: "Ask a Claude model twice for 30 test person names and the 2 lists share at least 1 name in 64 to 98% of pairs, depending on the model. Faker at the same size: 1.29%."

This is a mirror. The canonical home is the Zenodo record and the GitHub repository; the registration, prior work, limitations and full tables live there.

Paper: https://dimhold.by/papers/fixture-entropy/
Code and full write up: https://github.com/dimhold/fixture-entropy
DOI: https://doi.org/10.5281/zenodo.23545744

Files

path	what it is
`out/calls.jsonl`	every call: prompt, full reply, usage, timing (1366 rows: 1274 main, 66 n100, 26 failed and retried)
`out/baseline.json`	Faker, `uuid4` and `random` baseline values used in all tables
`out/summary-main.json`	per cell statistics from `report.py`
`out/report-main.md`	the same as tables
`out/extra-main.txt`	pooled rates, valid UUID recount, frequency lists, n100 pilot (`extra_stats.py`)
`out/logs/`	run logs
`context-check.json`	what the CLI could load from the environment

The names, emails, usernames, companies and addresses in the replies are invented by the models; they are the measured data, not real people.

How it was produced

1274 calls through Claude Code CLI 2.1.302 between 19 and 20 October 2026, each a separate process with no history, from an empty directory with tools, MCP servers and user settings disabled. 7 types, 3 arms (plain, json, diverse), 7 models. The model reported in modelUsage matched the one requested in all 1274. Rows with "ok": false are failed attempts that were retried. The n100 phase is incomplete and excluded from every conclusion. For publication the local user directory, the clean room path, the account email and a city in 1 reply were replaced with placeholders; nothing else in the journal was changed. Scripts that reproduce every number are in the GitHub repository.

Licence

MIT, code and data alike.

Citation

@misc{semenkevich2026fixtureentropy,
  author = {Semenkevich, Dmitriy},
  title  = {fixture-entropy: how often 2 LLM generated test fixtures collide},
  year   = {2026},
  doi    = {10.5281/zenodo.23545744},
  url    = {https://doi.org/10.5281/zenodo.23545744}
}