viking-13b-da-saga-delta-dpo

Viking-13B — Danish Grammar-Aligned (SAGA Δ-DPO)

Fine-tuned with SAGA (Syntax-Aware Grammar Alignment), a two-stage pipeline that trains language models to generate grammatically correct Danish text using reinforcement learning from a symbolic parser oracle (SpaCy (Danish dependency parser)).

This is a LoRA adapter. Load it on top of LumiOpen/Viking-13B.

SFT skipped (base PS 83.5% ≥ 80% threshold). Δ-DPO applied directly to base.

Results (independent Stanza evaluation)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("LumiOpen/Viking-13B", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "Hodfa71/viking-13b-da-saga-delta-dpo")

tokenizer = AutoTokenizer.from_pretrained("LumiOpen/Viking-13B")

prompt = "Dansk er"
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=60, temperature=0.8, do_sample=True)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Training details

• Data: 10 000 Danish Wikipedia sentences (filtered for quality)
• Method: Δ-DPO — generate N=8 candidates per prompt, keep pairs with parser score gap Δ ≥ 0.25, train with standard DPO loss (β=0.1)
• Parser oracle: SpaCy (Danish dependency parser)
• LoRA: rank 16, α=32, all linear layers, bfloat16
• Auto-SFT rule: SFT is applied first only if base parse success < 80%

Citation

@article{fakhar2025saga,
  title={SAGA: Syntax-Aware Grammar Alignment for Low-Resource Nordic Languages},
  author={Fakhar, Hoda and others},
  year={2025},
  note={Under review}
}

License

Inherits the base model license (AI Sweden LLM License / LumiOpen).