aimeri/name-generator

Name Generator

Ask for a name, get one name back and nothing else. A 3.4B fine-tune of Ministral 3 that answers requests like "a Brazilian man's name" or "a woman from the Ottoman Empire" with names drawn from real naming data, instead of the handful of names language models keep reaching for (Elara Voss, Kael Thorne, Dr. Sarah Chen).

v2 is out: aimeri/name-generator-v2 also invents fantasy names (elves, dwarves, orcs, dragons, fantasy kingdoms) and keeps the real ones.

GGUF quants (Q8_0, Q6_K, Q5_K_M, Q4_K_M) for llama.cpp, LM Studio and similar: aimeri/name-generator-GGUF

Examples

From the Q8_0 GGUF in llama.cpp: the greedy answer, then the first four distinct samples at temperature 0.8.

Request	Greedy	Sampled
What is a good male Brazilian name?	Lucas Silva	Alexandre Silva · Felipe Barbosa · Lucas Fernandes · Filipe Nogueira
Japanese woman's name	Kobayashi Masako	Maeda Yumi · Mizukami Masako · Kamimura Natsuko · Terada Shiori
German man born in the 1950s	Michael Schmidt	Gerhard Schülke · Jürgen Beyer · Michael Engel · Hermann Hoffmann
Give me a Russian man's full name	Vladimir Sergeyevich Kolesnikov	Maxwell Sergeevich Semyonov · Sergey Dmitriyevich Lebedev · Vasily Aleksandrovich Kolesnikov · Pavel Gusev
I need a name for a Mexican grandmother	María Teresa Martínez	Olivia García · María Díaz · María Teresa Díaz · Marina Rueda
A name from the Song dynasty	Zhu Xiu	Wei Xuan · Zhong Yuan · Zhao Xiu · Qian Zhong

Usage

Needs transformers 5.x (the architecture is Ministral3ForCausalLM).

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "aimeri/name-generator"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")

def names(request, n=5, temperature=0.8):
    text = tokenizer.apply_chat_template([{"role": "user", "content": request}],
                                         tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(text, return_tensors="pt").to(model.device)  # adds <s>, as in training
    out = model.generate(**inputs, max_new_tokens=24, do_sample=True, temperature=temperature,
                         num_return_sequences=n)
    start = inputs["input_ids"].shape[1]
    return [tokenizer.decode(o[start:], skip_special_tokens=True).strip() for o in out]

print(names("Give me a Brazilian woman's name"))

Two details matter:

• Render, then tokenize. The ChatML template emits no <s>; the tokenizer adds it when it encodes the rendered text. apply_chat_template(..., tokenize=True) skips it, which is not what the model was trained on.
• Plain-string content. The template trims content as a string; a list of {"type": "text"} parts ends up in the prompt as its Python repr.

Through an OpenAI-compatible server (llama.cpp, vLLM, LM Studio) send a single user message with max_tokens around 24. The model ends its turn with <|im_end|> (the eos token).

Getting good names

• Name a country or nationality ("Nigerian", "from Peru"), a historical state ("Ottoman Empire", "Soviet Union", "Song dynasty") or a region ("West African").
• Add sex and era for more control: "a woman born in the 1950s" shifts the first names to that decade.
• Names follow local conventions: family name first for Chinese, Japanese, Korean, Hungarian and Vietnamese; patronymics in Russian and Ukrainian full names; two surnames in Spanish and Portuguese full names.
• Sample at temperature 0.7–1.0 and ask once per character. Greedy decoding returns the single most typical name every time.

Training

Base	aimeri/Ministral-3-3B-Base-2512-chatml: the text-only Ministral-3-3B-Base-2512 with `<
Data	aimeri/name-generator, tag v1: 8,536 training and 459 validation examples covering 333 countries and historical states (24% historical)
Method	Supervised fine-tuning with Unsloth and TRL, full fine-tune, loss on the assistant turn only
Hyperparameters	learning rate 1e-5, linear schedule, 5% warmup, 1 epoch (2,134 steps), effective batch size 4, AdamW 8-bit, weight decay 0.001
Software	transformers 5.20.2, Unsloth 2026.10.9

Each example pairs a request with one name. The names were sampled from real frequency data (national statistics offices such as the US SSA and Census, INSEE, ONS, INE, IBGE, PESEL, SSB and SCB; Wikidata people by citizenship and birth decade; Faker and Mimesis locale lists), by country, sex and birth decade. Ministral-3-3B-Instruct-2512 wrote the requests from each name's metadata. See the dataset card for details.

Limitations

• Real-world names only. There were no fantasy names in training. Asked for an elf name, it returns English-sounding names ("Lily Lore", "Zoe Lightfoot").
• Always a full name. The training data had no first-name-only or surname-only answers, so "a Nigerian surname" still gets a full name.
• Sex can slip when sampling: asked for a woman's name from the Ottoman Empire, some samples are men's names.
• Coverage follows the data. Where a country has no official name statistics, names come from Wikidata's notable people, who skew older, male and elite. Some pairings reflect expatriates or data errors (a "Maxwell Sergeevich" among the Russian names).
• Short English requests. Training used single-turn requests of about one sentence.
• Not real people. Outputs are combinations of real given names and surnames. Any match with a real person is coincidental. Don't use this to identify people, infer anyone's background, or make documents.

License

Apache 2.0, following the base model. The training dataset is CC BY-SA 4.0.