lapa-instruct-uk-en-grpo

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

Trained with GRPO using Condition A: chrF + BLEU lexical rewards only.

Results

Benchmark	BLEU	chrF	CometKiwi	xCOMET	Mistral Judge	Aya Judge
FLoRes+ devtest	43.15	67.85	0.7658	0.9473	88.62	94.33
WMT24	35.75	60.70	0.7053	0.8212	84.74	91.84

Baseline (lapa-v0.1.2-instruct): 42.02 FLoRes BLEU / 34.60 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-uk-en-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 (0.55), BLEU (0.45)
• Data: WikiMatrix uk-en (132K pairs), 300 steps
• Infrastructure: DeepSpeed ZeRO-2, 4x RTX 6000 Ada

See reward-driven-translation for full reproduction code.