from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.2-1B",
    dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "Hodfa71/llama-3.2-1b-nb-saga-delta-dpo")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")

inputs = tokenizer("Regjeringen kunngjorde i dag", return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=80, do_sample=True, temperature=0.8)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Training

Method: Δ-DPO (online preference pairs filtered by quality gap δ≥0.25)
Grammar oracle: SpaCy nb_core_news_lg
SFT: skipped (base PS = 83% > τ=0.80)
Antihack: repetition_penalty=1.3, mattr_window=20
LoRA: r=16, α=32, all-linear target modules, bfloat16
Training data: Norwegian Bokmål Wikipedia prompts

Citation

bibtex

@article{fakhar2025saga,
  title   = {SAGA: Syntax-Aligned Grammar Adaptation via Reinforcement Learning from Verifiable Feedback},
  author  = {Fakhar, Hoda and Carlsson Wiman, Emil},
  year    = {2025}
}