Shan Gemma 4 E2B LoRA is an instruction-tuned Shan language model built upon Google's Gemma 4 architecture.
The model was trained on a custom Shan educational instruction dataset.
1ၼမ်ႉ ၼႆႉ မၢႆထိုင် water။
2တူဝ်ယၢင်ႇ - ၵဝ်ၶႃႈ ၶႂ်ႈၵိၼ်ၼမ်ႉ။
1Translate to Shan:
2
3I want to learn Shan language.
1ၼၢမ်ႇၼႆႉ ပဵၼ်ၶေႃႈၵႂၢမ်းဢၼ်ၸီႉၸူး
2ၵူၼ်း၊ သတ်း၊ တီႈယူႇ ဢမ်ႇၼၼ်
3သင်ႇၶွင်ႇတၢင်းမူတ်း။
1from unsloth import FastLanguageModel
2
3model, tokenizer = FastLanguageModel.from_pretrained(
4 model_name="k1ngtai/shan-gemma4-E2B-it-lora",
5 max_seq_length=1024,
6 load_in_4bit=True,
7)
8
9FastLanguageModel.for_inference(model)
10
11prompt = """<start_of_turn>user
12You are a Shan language learning assistant.
13
14Instruction: Teach Shan vocabulary.
15Input: Give me 5 beginner Shan words about family.
16<end_of_turn>
17<start_of_turn>model
18"""
19
20inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
21
22outputs = model.generate(
23 **inputs,
24 max_new_tokens=300,
25 temperature=0.7,
26 top_p=0.9,
27 do_sample=True,
28)
29
30print(tokenizer.decode(outputs[0], skip_special_tokens=True))
This model is intended for peaceful educational use only.
It is focused on Shan/Tai language learning and community education.
1@misc{shan-gemma4-e2b-lora,
2 title={Shan Gemma 4 E2B LoRA},
3 author={K1ngtai},
4 year={2026},
5 publisher={Hugging Face}
6}
