language, license_name, license_link, tags, base_model, pipeline_tag

language	license_name	license_link	tags	base_model	pipeline_tag
en uk	lfm1.0	https://www.liquid.ai/lfm-license	translation grpo ukrainian lfm2	Yehor/kulyk-en-uk	text-generation

kulyk-en-uk-grpo

English-to-Ukrainian translation model based on Yehor/kulyk-en-uk (LFM2-350M), improved with GRPO using calibrated guardrail rewards on WikiMatrix data.

Results

FLoRes+ devtest (sentence-by-sentence, greedy, repetition_penalty=1.05)

Model	BLEU	chrF	CometKiwi
kulyk-en-uk (baseline)	27.28	57.06	0.7228
kulyk-en-uk-grpo	27.83	57.42	0.7171

WMT24 en-uk (news domain, out-of-distribution)

Model	BLEU	chrF	CometKiwi
kulyk-en-uk (baseline)	21.57	48.82	0.5633
kulyk-en-uk-grpo	23.23	49.80	0.5640

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("iamthewalrus67/kulyk-en-uk-grpo", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("iamthewalrus67/kulyk-en-uk-grpo", trust_remote_code=True)

prompt = "Translate the text to Ukrainian:\nThe weather is beautiful today."
input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True, return_tensors="pt", tokenize=True
).to(model.device)

output = model.generate(input_ids, max_new_tokens=256, do_sample=False, repetition_penalty=1.05)
print(tokenizer.decode(output[0][input_ids.shape[1]:], skip_special_tokens=True))

Training Details

• Method: GRPO with calibrated guardrail rewards
• Data: WikiMatrix en-uk (132K pairs)
• Rewards: chrF (0.30) + BLEU (0.25) + CometKiwi (0.20) + 5x calibrated guardrails (0.05 each)
• Training: Full fine-tune (no LoRA), single GPU, 300 steps
• Best checkpoint: step 100

See reward-driven-translation for full reproduction code.