Fine-tuned with SAGA (Syntax-Aware Grammar Alignment). Danish base PS (80.5%) is at the τ=0.80 threshold, so SFT is skipped and Δ-DPO is applied directly from base.

This is a LoRA adapter. Load it on top of meta-llama/Llama-3.1-8B.

Note: Llama 3.1 8B is a gated model — request access at HuggingFace before downloading.

Results (Stanza DA — held-out evaluator)

Metric	Base	+ Δ-DPO (no SFT)
Stanza PS ↑	80.5%	94.0%
Stanza score ↑	0.388	0.592
PPL-Wiki ↓	13.2	13.2
Summarisation PS ↑	50.0%	92.5%

+13.5pp parse success with no PPL degradation. No-SFT Δ-DPO outperforms every KL-SFT variant (91.0–91.5%), confirming the τ=0.80 Auto-SFT skip rule.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "Hodfa71/llama-3.1-8b-da-saga-delta-dpo")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B")

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

Base model: Llama 3.1 8B (base DA PS 80.5% ≥ τ=0.80 → SFT skipped)
Stage: Δ-DPO directly from base — N=8 candidates, δ≥0.25, β=0.1
Oracle: SpaCy da_core_news_lg (Danish dependency parser)
LoRA: rank 16, α=32, all linear layers, bfloat16

Citation

bibtex

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

License

Meta Llama 3.1 Community License.