danskgpt-v4-12b

danskgpt-v4-12b

A Danish fine-tune of google/gemma-4-12B-it, tuned to be strong on Danish understanding, knowledge and reasoning tasks. It is a text-only Gemma4ForCausalLM checkpoint in bf16. For an even stronger model see syvai/danskgpt-v4-31b.

Usage

The model uses Gemma-4's chat format — always go through apply_chat_template (or vLLM's chat) so the <|turn> / <turn|> markers are added correctly.

Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "syvai/danskgpt-v4-12b"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")

messages = [
    {"role": "user", "content": "Forklar kort, hvad fotosyntese er, og hvorfor det er vigtigt."},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.9)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

vLLM

from vllm import LLM, SamplingParams

llm = LLM(model="syvai/danskgpt-v4-12b", dtype="bfloat16")
messages = [
    {"role": "user", "content": "Giv mig tre forslag til en hyggelig weekendtur i Jylland."},
]
out = llm.chat(messages, SamplingParams(temperature=0.7, top_p=0.9, max_tokens=512))
print(out[0].outputs[0].text)

More Danish prompt ideas: "Omskriv denne sætning, så den bliver mere formel: ...", "Opsummer teksten nedenfor i tre punkter på dansk: ...", "Er følgende anmeldelse positiv eller negativ? ...".

Training

danskgpt-v4-12b is a QLoRA fine-tune (rank 16, 4-bit NF4 base, bf16 compute) of google/gemma-4-12B-it, trained on Danish task data to strengthen Danish understanding, knowledge and reasoning.

Intended use & limitations

• Intended use: Danish NLP — classification, reading comprehension, knowledge/reasoning Q&A, NER, and general Danish assistant tasks.
• Limitations: As a task-tuned model, it can be more terse/decisive than the base on open-ended generation. It inherits the base model's biases and knowledge cutoff. Not intended for high-stakes use without human oversight.

License

Released under CC-BY-4.0, consistent with other SyvAI models. Use of the base model is additionally subject to Google's Gemma terms.