AgroNVILA (GRPO)

Final checkpoint of AgroNVILA: supervised fine-tuning on AgroOmni, then Group Relative Policy Optimization (GRPO) alignment.

Training

• Corpus: AgroOmni, 268,119 question-answer pairs over 97,926 images
• Stage 1: supervised fine-tuning, 59.52 overall on AgroMind
• Stage 2: GRPO with task-specific rewards, 59.86 overall
• The SFT-only checkpoint is AgroOmni/AgroNVILA-SFT
• Code: https://github.com/Jrookie2005/AgroOmni

Results

AgroMind benchmark, 13 tasks: 59.86 overall micro-average, the SFT baseline reaches 59.52. Anomaly Reasoning 75.74, Boundary Detection 58.02, Area Statistics 70.18, Visual Prompt Reasoning 55.63.

Usage

from llava.model.builder import load_pretrained_model
tokenizer, model, image_processor, _ = load_pretrained_model(
    "AgroOmni/AgroNVILA", None, "llava_llama"
)

License

Apache-2.0, inherited from the NVILA-Lite-8B base model.