You can deploy this model and start earning money today!

emilcw/
gemma4-e4b-nb-saga-kl-sft-delta-dpo

Name: gemma4-e4b-nb-saga-kl-sft-delta-dpo
Availability: InStock

Views

No views yet

Gemma4-E4B — Norwegian Bokmål Grammar-Aligned (SAGA Δ-DPO)

Fine-tuned with SAGA (Syntax-Aware Grammar Alignment), a two-stage pipeline that trains language models to generate grammatically correct Norwegian Bokmål text using reinforcement learning from a symbolic parser oracle (SpaCy nb_core_news_lg (Norwegian dependency parser)).

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

KL-SFT + Δ-DPO (non-Nordic control model, ablation condition). Base PS 84.5% ≥ τ=0.80 — SFT normally skipped; forced for ablation. 100% parse success post-DPO. No-SFT variant is recommended. SpaCy nb_core_news_lg oracle. Held-out eval: Stanza nb. ScaLA NB AUROC: 0.593 → 0.590. Summarisation PS: 44.5% → 100%.

Results (independent Stanza evaluation)

Metric	Base	+ Δ-DPO
Parse success	84.5%	100.0%
Parse score	0.391	0.666
PPL-Wiki	16.8	29.6

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("google/gemma-4-E4B", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "emilcw/gemma4-e4b-nb-saga-kl-sft-delta-dpo")

tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-E4B")

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

Data: 10 000 Norwegian Bokmål Wikipedia sentences (filtered for quality)
Method: Δ-DPO — generate N=8 candidates per prompt, keep pairs with parser score gap Δ ≥ 0.25, train with standard DPO loss (β=0.1)
Parser oracle: SpaCy nb_core_news_lg (Norwegian dependency parser)
LoRA: rank 16, α=32, all linear layers, bfloat16
Auto-SFT rule: SFT is applied first only if base parse success < 80%

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

Inherits the base model license (AI Sweden LLM License / LumiOpen).