# Earthmind-R1

EarthMind-4B fine-tuned with GRPO (Group Relative Policy Optimization) for geospatial visual question answering.

## Model Details

- **Base Model**: EarthMind-4B (InternVL-based architecture)
- **Training Method**: GRPO with LoRA adapters
- **Training Data**: Geospatial instruction dataset
- **Output Format**: Chain-of-thought reasoning with `<think>` and `<answer>` tags

## Usage

```python
import torch
from transformers import AutoModel, AutoTokenizer
from PIL import Image

model = AutoModel.from_pretrained(
    "aadex/Earthmind-R1",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("aadex/Earthmind-R1", trust_remote_code=True)

# Prepare for generation
model.preparing_for_generation(tokenizer=tokenizer, max_new_tokens=512, torch_dtype=torch.bfloat16)

# Load your image
image = Image.open("your_image.jpg")

# Create prompt
question = "Describe what you see in this satellite image."
prompt = f"""User: <image>
{question} First output the thinking process in <think> </think> tags and then output the final answer in <answer> </answer> tags.
Assistant:"""

# Generate (use model's chat method or manual generation)
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(response)
```

## Training

Trained using GRPO with:
- LoRA rank: 16
- LoRA alpha: 32
- Learning rate: 5e-6
- Epochs: 3
- Reward functions: accuracy, format

## License

Please refer to the base EarthMind-4B model license.