emilcw/llama-3.2-1b-nb-saga-kl-sft-delta-dpo

Llama-3.2-1B — Norwegian Bokmål Grammar-Aligned (SAGA KL-SFT + Δ-DPO) [Ablation]

Fine-tuned with SAGA (Syntax-Aware Grammar Alignment) using forced KL-regularised SFT followed by Δ-DPO. This is an ablation model — the base model's parse success (83%) already exceeds the τ=0.80 Auto-SFT threshold, so SAGA would normally skip SFT. This model forces KL-SFT to measure the effect of the extra SFT stage.

This is a LoRA adapter. Load it on top of meta-llama/Llama-3.2-1B.

Ablation finding: Forcing KL-SFT when base PS ≥ 80% does not improve results. Parse success is comparable (98.0% vs 98.5%) but PPL degrades (+10.8) and ScaLA AUROC drops (0.592 vs 0.617). The no-SFT variant (emilcw/llama-3.2-1b-nb-saga-delta-dpo) is the recommended model.

Results (Stanza NB — independent held-out evaluator)

Metric	Base	No-SFT Δ-DPO	KL-SFT + Δ-DPO (this)
Stanza PS ↑	83.0%	98.5%	98.0%
Parse score ↑	0.392	0.622	0.630
PPL-Wiki ↓	30.1	33.0	43.8
ScaLA AUROC ↑	0.605	0.617	0.592
Summ PS ↑	63.0%	100%	100%
Summ score ↑	0.202	0.722	0.758
RC APS ↑	31.0%	87.5%	93.0%

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "emilcw/llama-3.2-1b-nb-saga-kl-sft-delta-dpo")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")

prompt = "Norsk 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 3.2 1B (base NB PS 83% ≥ τ=0.80, but SFT forced for ablation)
• Stage 1: KL-SFT on 10k Norwegian Wikipedia sentences, 3 epochs, λ=0.10
• Stage 2: Δ-DPO from KL-SFT checkpoint — N=8 candidates, δ≥0.25, β=0.1
• Anti-hacking: MATTR diversity weight=0.2, repetition_penalty=1.3
• Oracle: SpaCy nb_core_news_lg (Norwegian dependency parser)
• LoRA: rank 16, α=32, all linear layers, bfloat16

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

Meta Llama 3.2 Community License.