Fine-tuned with SAGA (Syntax-Aware Grammar Alignment) using KL-regularised SFT followed by Δ-DPO. The KL penalty keeps the policy anchored to the base distribution during SFT, preserving generation diversity for downstream preference mining.

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 — you need to request access at HuggingFace before downloading.

Results (Greynir IS constituency parser — primary oracle for Icelandic)

Model	Greynir PS ↑	Score ↑	PPL-Wiki ↓	ScaLA AUROC ↑
Llama-3.1-8B Base	72.0%	0.443	22.0	0.583
+ KL-SFT	63.5%	0.502	14.1	—
+ KL-SFT + Δ-DPO	96.0%	0.703	17.1	0.615

+24pp Greynir parse success over base. Best IS result for a non-Nordic model.

Greynir is a rule-based Icelandic constituency parser (purpose-built for Icelandic morphology). Stanza IS is used as an independent held-out eval but is weaker for Icelandic.

Auto-SFT rule (§9.1)

Llama 3.1 8B has no Icelandic/Nordic training data (base PS 72% < τ=80%, non-Nordic). The three-way rule applies:

PS ≥ 0.80 → skip SFT
PS < 0.80, Nordic pretrain → standard SFT
PS < 0.80, non-Nordic → KL-SFT ← this model

Standard 3-epoch SFT collapses entropy (2.70 → 2.10), starving Δ-DPO of diverse pairs. KL-SFT (λ=0.10) retains entropy and raises PS to 96%.

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-is-saga-kl-sft-delta-dpo")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B")

prompt = "Íslenska 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 (no Nordic pretraining)
Stage 1: KL-SFT on 10k Icelandic Wikipedia sentences, 3 epochs, λ=0.10
Stage 2: Δ-DPO from merged KL-SFT model — N=4 candidates, δ≥0.25, β=0.1
Oracle: Greynir (Icelandic constituency parser)
LoRA: rank 16, α=32, all linear layers (q/k/v/o/gate/up/down), 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. The LoRA adapter weights are released under the same license as the base model.