Blame for README.md

---
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: domain
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: name
    dtype: string
  - name: sex
    dtype: string
  - name: decade
    dtype: int64
  - name: country
    dtype: string
  - name: country_label
    dtype: string
  - name: region
    dtype: string
  - name: language
    dtype: 'null'
  - name: historical
    dtype: bool
  - name: archetype
    dtype: string
  - name: kind
    dtype: string
  - name: flavor
    dtype: string
  - name: setting
    dtype: string
  - name: role
    dtype: string
  - name: tone
    dtype: string
  - name: culture
    dtype: string
  splits:
  - name: train
    num_bytes: 3137542
    num_examples: 13217
  - name: validation
    num_bytes: 167677
    num_examples: 708
  download_size: 1007355
  dataset_size: 3305219
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
language:
- en
license: cc-by-sa-4.0
size_categories:
- 10K<n<100K
task_categories:
- text-generation
pretty_name: Name Generator
tags:
- names
- character-names
- fantasy-names
- creative-writing
- synthetic
- sft
---

# Name Generator

Short requests for a character name, each answered with a name and nothing else: real-world names
for real-world requests ("a Brazilian man's name" → *Paulo Rozeno Freitas*) and invented names for
fantasy requests ("a grim name for a dwarven blacksmith"). Made for fine-tuning small models out of
the few names language models keep reaching for (Elara Voss, Kael Thorne, Dr. Sarah Chen).

| split | rows | real | fantasy |
|---|---|---|---|
| train | 13,217 | 8,536 | 4,681 |
| validation | 708 | 459 | 249 |

```json
{"domain": "real", "messages": [{"role": "user", "content": "What is a Spanish-inspired male name with a Spanish-sounding surname?"}, {"role": "assistant", "content": "David Úbeda"}]}
{"domain": "fantasy", "messages": [{"role": "user", "content": "A harsh but evocative name for a grizzled hill dwarf smith in my grim mountain kingdom"}, {"role": "assistant", "content": "Korgell Gellbau"}]}
```

The `v1` tag holds the earlier real-names-only release.

## How it was made

### Real names

1. **Names.** Full names were sampled with namebank from public name data: national statistics
   offices (US SSA and Census, INSEE, ONS, INE, IBGE, PESEL, SSB, SCB and others), Wikidata people
   grouped by citizenship and birth decade, and Faker/Mimesis locale lists. Names follow real
   frequency distributions by country, sex and birth decade; first names and surnames are paired
   the way they co-occur; local conventions apply (name order, double surnames, gendered surnames,
   patronymics). Coverage: 333 countries and historical states; 2,129 rows
   (24%) from historical states. Most frequent: China (330), India (329), United States (166), Indonesia (147), Pakistan (137), Nigeria (134), Brazil (128), Bangladesh (128).
2. **Requests.** Each name, with its sex, birth decade and country, went to
   Ministral-3-3B-Instruct-2512 (Q8_0 GGUF), which wrote a request that the name answers.

### Fantasy names

1. **Names.** `namebank fantasy` invents them. Species (elves, dwarves, orcs, dragons, fae...) get
   invented naming languages, each culture with its own sounds, syllable shapes and morphemes
   drawn from a palette for its species; species never draw on real languages. Human cultures
   are modeled on real naming traditions (Anglo-Saxon, Arabic, Caucasian, Celtic, Chinese, East African, Finnic, Frankish, Germanic, Hellenic, Japanese, Latin, Magyar, Norse, Persian, Polynesian, Slavic, South Asian, Turkic, West African) through culture-wide sound changes and a
   character model trained on the changed names. Halflings get compound family names from homely
   words. A tone (grim, gentle, whimsical, noble...) tilts each name's sounds toward it.
2. **Guarantees.** No given name or family name is a real name in namebank's data (every name from
   the sources above), none is on a list of names language models overuse or a list of famous
   fictional names, or within about one edit of either, and none contains an obscenity.
3. **Requests.** The same model wrote a request from each name's kind (elf, dwarf, human...), role,
   tone and, for human cultures, the tradition it borrows from; a random share of details was left
   out of each prompt. Requests for human characters always signal a fantasy setting, so they
   don't read as real-name requests. By archetype: human-inspired (1,184), human (756), dwarf (447), elf (420), orc (296), fae (218), vampire (205), goblin (199), gnome (197), demon (196), halfling (186), dragon (175), giant (118), merfolk (110), lizardfolk (96), dark elf (70), eldritch (57).

### Cleanup

Requests that contained the answer, named another character, ran over several lines or (for
fantasy humans) gave no fantasy cue were dropped (2,075 rows).

## Fields

- `messages`: the request (`user`) and the answer (`assistant`, the name alone), ready for chat SFT.
- `domain`: `real` or `fantasy`. `name`: the answer again. `sex`: `M`, `F` or `X` (unisex, or sex
  doesn't apply).
- Real names: `decade` (birth decade), `country` (ISO code, or a Wikidata QID for historical
  states), `country_label`, `region` (UN subregion), `language`, `historical`.
- Fantasy names: `archetype` (elf, dwarf, human-inspired...), `kind` (what the request writer was
  told), `flavor` (the real tradition a human culture borrows from), `setting`, `role`, `tone`,
  `culture` (names sharing it come from one invented culture).

## Splits

`validation` holds about 5% of the rows, picked by a hash of the answer, so no name
appears in both splits.

## Use

Supervised fine-tuning of a name generator. Compute the loss on the assistant turn only: the answer
is a handful of tokens, and the requests were written by a small model.

## Limitations

- Wikidata covers notable people, who skew older, male and elite. Countries without official name
  statistics inherit that skew.
- Requests come from a 3B model: phrasing repeats, and some are vaguer than the facts behind them.
- A real full name here is a sampled combination of real first names and surnames, not a record of
  a real person; any match with a real person is coincidental.
- Fantasy names are generated, not curated: some are clumsy, and cultures modeled on real traditions
  reflect the generator's sound changes rather than those languages.
- Spelling and romanization follow each source.

## License

CC BY-SA 4.0, because some of the name lists namebank draws on are share-alike (IBGE's 2010
census via brasil.io, Wiktionary, JMnedict). Per-source licenses, URLs and retrieval dates are in [SOURCES.md](SOURCES.md).