Hodfa71/gemma4-e4b-is-saga-kl-sft-delta-dpo

Gemma4-E4B — Icelandic Grammar-Aligned (SAGA KL-SFT + Δ-DPO + MATTR)

Fine-tuned with SAGA (Syntax-Aware Grammar Alignment) using KL-regularised SFT followed by Δ-DPO with MATTR diversity penalty. Non-Nordic model with no Icelandic pretraining — standard SFT collapses distribution; KL-SFT + MATTR prevents hacking.

This is a LoRA adapter. Load it on top of google/gemma-4-E4B-it.

Results (Stanza IS — independent held-out evaluator)

Metric	Base	No-SFT Δ-DPO	KL-SFT + Δ-DPO + MATTR
Stanza PS ↑	78.0%	80.0%	83.5%
Stanza score ↑	0.351	0.314	0.379
MATTR ↑	0.795	0.766	0.919
PPL-Wiki ↓	12.2	—	13.5

Greynir (oracle) PS rises from 25.5% (base) to 86.0% with KL-SFT + Δ-DPO + MATTR. Standard SFT drops Stanza PS by 9.5pp and nearly doubles PPL — KL-SFT prevents both.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("google/gemma-4-E4B-it", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "Hodfa71/gemma4-e4b-is-saga-kl-sft-delta-dpo")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-E4B-it")

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: Gemma 4 E4B (no Nordic pretraining; base IS PS 78% < τ=80%)
• Stage 1: KL-SFT on 10k Icelandic Wikipedia sentences, 3 epochs, λ=0.10
• Stage 2: Δ-DPO from merged KL-SFT model — N=8 candidates, δ≥0.25, β=0.1
• Anti-hacking: MATTR diversity weight=0.2, repetition_penalty=1.3
• Oracle: Greynir (Icelandic constituency 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}
}