Galactus

This model is a fine-tuned version of microsoft/Phi-4-multimodal-instruct on the Galaxy's Last Exam Benchmark.

Model description

Galactus is a SOTA multimodal language model that outperforms all OpenAI and Gemini models on the Galaxy's Last Exam Benchmark. This benchmark features challenging tasks that push the boundaries of metaphysical competence—for instance, determining how many times two lines intersect or simulating the effect of adding three minutes to an analog clock. The model accepts image input along with text prompts and has been specifically optimized to tackle the most complex visual reasoning tasks.

Intended uses & limitations

This model is intended for handling complex visual reasoning tasks that require metaphysical competence. Please do not use for normal human tasks.

Training and evaluation data

The model was exclusively trained on the Galaxy's Last Exam Benchmark.

Training procedure

The model was trained using LoRA adapters focused on the vision components of the base model.

Prompt format

This model uses the following image prompt format: <|image_1|> + user text

Training hyperparameters

The following hyperparameters were used during training:

num_train_epochs: specified in args (but used checkpoint @ 252 epochs) per_device_train_batch_size: specified in args gradient_checkpointing: True gradient_checkpointing_kwargs: {'use_reentrant': False} gradient_accumulation_steps: specified in args optim: 'adamw_torch' adam_beta1: 0.9 adam_beta2: 0.95 adam_epsilon: 1e-7 learning_rate: specified in args weight_decay: 0.0 save_strategy: 'steps' save_steps: 10 eval_steps: 10 if eval_dataset else None evaluation_strategy: 'steps' if eval_dataset else 'no' load_best_model_at_end: True if eval_dataset else False max_grad_norm: 1.0 lr_scheduler_type: 'linear' warmup_steps: 50 logging_steps: 10 save_total_limit: 2 save_only_model: True dataloader_num_workers: 4 ddp_find_unused_parameters: True

Training results

The model achieved 72% performance on the Galaxy's Last Exam Benchmark. https://github.com/DavidTee1/Galaxys-Last-Exam-Benchmark

Framework versions

Transformers 4.46.1 PyTorch 2.7.0.dev20250304+cu128 TorchVision 0.22.0.dev20250304+cu128 Tokenizers 0.20.3