from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model
model = AutoModelForCausalLM.from_pretrained(
    "bhismaperkasa/gemma-3-270m-form-generator-bf16",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("bhismaperkasa/gemma-3-270m-form-generator-bf16")

# Generate form
messages = [
    {"role": "system", "content": "You are a helpful assistant that generates form definitions in JSON format."},
    {"role": "user", "content": "buatkan form login dengan email dan password"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True
)

result = tokenizer.decode(outputs[0], skip_special_tokens=True)
generated = result.split("<start_of_turn>model\n")[-1].strip()
print(generated)

For Android (MediaPipe)

Step 1: Convert to TFLite

bash

# Clone training repo
git clone https://huggingface.co/bhismaperkasa/gemma-3-270m-form-generator-bf16
cd gemma-3-270m-form-generator-bf16

# Convert (requires ai-edge-torch)
python convert_to_tflite.py --model_path ./

Step 2: Integrate to Android

See full Android integration guide in training repository.

📱 Android Performance

Device	Init Time	Inference Time	Memory
Flagship (2024)	2-3s	1-2s	~200MB
Mid-range (2023)	3-5s	2-4s	~200MB
Budget (2022)	5-8s	4-8s	~250MB

📋 Example Outputs

Input

buatkan form pendaftaran event dengan nama, email, dan nomor telepon

Output

json

{
  "id": "form_event_registration_001",
  "title": "Form Pendaftaran Event",
  "category": "registration",
  "formDefinition": {
    "sections": [
      {
        "sectionId": "section_1",
        "title": "Informasi Peserta",
        "fields": [
          {
            "fieldId": "nama_lengkap",
            "label": "Nama Lengkap",
            "fieldType": "TEXT",
            "required": true
          },
          {
            "fieldId": "email",
            "label": "Email",
            "fieldType": "EMAIL",
            "required": true
          },
          {
            "fieldId": "nomor_telepon",
            "label": "Nomor Telepon",
            "fieldType": "PHONE",
            "required": true
          }
        ]
      }
    ]
  }
}

🔧 Technical Notes

Why NO Quantization?

Model ini di-train tanpa quantization (4-bit/8-bit) karena:

TFLite Compatibility: Quantized model sulit di-convert ke TFLite
Merge Quality: Quantized LoRA merge sering corrupt output
Android Optimization: TFLite akan apply quantization sendiri (int8)

Model Size

PyTorch (bfloat16): ~500MB
TFLite (float32): ~250MB
TFLite (int8): ~130MB ⭐ Recommended untuk Android

📚 Dataset

Trained on bhismaperkasa/form_dinamis

Size: 10,000 form examples
Language: Bahasa Indonesia
Format: Conversational (messages format)
Split: 90% train / 10% validation

🎓 Training Repository

Full training code, conversion scripts, and Android integration guide:

Training scripts (nohup support for RunPod)
TFLite conversion
Android integration examples
Complete documentation

⚖️ License

Apache 2.0 (following Gemma base model license)

🤝 Citation

bibtex

@misc{form-generator-gemma-android,
  author = {bhismaperkasa},
  title = {Form Generator - Gemma 3 270M (Android-Ready)},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/bhismaperkasa/gemma-3-270m-form-generator-bf16}}
}

📞 Support

For issues, questions, or contributions, please visit the training repository or open an issue.

Ready for production Android deployment! 🚀📱