lapa-instruct-en-uk-grpo

English-to-Ukrainian translation LoRA adapter for lapa-v0.1.2-instruct (Gemma-3 12B).

Trained with GRPO using Condition E: LLM judge (M-Prometheus-14B) + calibrated guardrail rewards.

Results

Benchmark	BLEU	chrF	CometKiwi	xCOMET	Mistral Judge	Aya Judge
FLoRes+ devtest	34.02	61.63	0.7889	0.9357	83.94	92.00
WMT24	31.24	57.19	0.6930	0.7995	80.44	89.11

Baseline (lapa-v0.1.2-instruct): 33.44 FLoRes BLEU / 31.12 WMT24 BLEU.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("lapa-llm/lapa-v0.1.2-instruct", device_map="auto", dtype="bfloat16")
model = PeftModel.from_pretrained(base_model, "iamthewalrus67/lapa-instruct-en-uk-grpo")
tokenizer = AutoTokenizer.from_pretrained("lapa-llm/lapa-v0.1.2-instruct")

Training Details

• LoRA: r=128, alpha=256, targets=q/k/v/o projections
• Rewards: chrF (0.10), BLEU (0.05), LLM judge (0.60), 5x calibrated guardrails (0.05 each)
• Data: WikiMatrix en-uk (132K pairs), 300 steps
• Infrastructure: DeepSpeed ZeRO-2, 4x RTX 6000 Ada

See reward-driven-translation for full reproduction code.