RinKana/arc_grid_relational-model

🧩 ARC Grid Relational Model

Author: RinKana
License: Apache 2.0
Framework: TensorFlow / Keras
Task: Abstract Reasoning — Grid-to-Grid Transformation (ARC)
Status: Research / Prototype

---

🧠 Model Summary

arc_grid_relational-model is a grid-relational transformer designed for the Abstraction and Reasoning Corpus (ARC) tasks.
It learns to predict output grids from input grids by combining local convolutional encoding with global relational self-attention, allowing reasoning over both spatial and relational features.

Architecture Overview

• Encoder: Two convolutional layers with GELU + normalization
• Projection: 1×1 convolution to per-cell embedding vectors
• Relational Transformer: Multi-layer self-attention over flattened grid cells
• Readout Head: MLP → color logits per cell
• Loss: Masked Sparse Categorical Cross-Entropy
• Metric: Masked Accuracy (ignores padded cells)

This design aims to encourage pattern discovery and relational reasoning in a grid-based environment similar to the ARC benchmark.

---

🚀 Intended Use

Use Cases

• Research on visual reasoning and combinatorial generalization
• Exploring relational inductive biases in neural architectures
• As a base encoder for hybrid symbolic–neural ARC solvers

Limitations

• Not fine-tuned for every ARC task — generalization is limited
• Does not perform explicit symbolic reasoning or program induction
• May fail on tasks requiring hierarchical or multi-step reasoning
• Sensitive to grid size beyond the training maximum

---

📊 Training Details

Setting	Value
Dataset	ARC Prize 2025 training challenges (arc-agi_training_challenges.json)
Grid Input	Integer-encoded colors (0–9), padded to max H×W
Input Encoding	One-hot grid → (H, W, num_colors)
Optimizer	Adam (lr = 3e-4)
Batch Size	32
Epochs	10
Random Seed	42

Model Hyperparameters

Parameter	Value
Conv Channels	[64, 128]
Cell Embedding Dim	192
Transformer Layers	4
Attention Heads	8
FFN Dim	512
Dropout	0.1

Evaluation Metric

• Masked Accuracy: Average per-cell correctness ignoring -1 (padding mask)
• (Optionally) full-grid exact match rate for reasoning-level validation

---

🧩 Model Architecture Diagram (Conceptual)

Input Grid → Conv Encoder → Flatten to Sequence ↓ Relational Transformer ↓ Per-cell MLP Readout → Output Grid (Color logits)

---

🧪 Example Usage

import tensorflow as tf
import numpy as np

model = tf.keras.models.load_model("path/to/arc_grid_relational_model")

# Example input grid (integer colors)
inp_grid = np.array([
    [1, 1, 0],
    [0, 2, 2],
    [0, 0, 0]
], dtype=np.int32)

# Pad + one-hot encode
H, W = inp_grid.shape
H_max, W_max = 30, 30   # example maximum grid size used during training
num_colors = 10

padded = np.zeros((H_max, W_max), dtype=np.int32)
padded[:H, :W] = inp_grid
x = tf.one_hot(padded, depth=num_colors)
logits = model(x[None, ...])
pred = tf.argmax(logits, axis=-1)[0][:H, :W].numpy()

print(pred)

---

⚖️ License

Licensed under the Apache License 2.0. You may use, modify, and distribute this model under the same terms.

🧬 Citation

If you use this model in your research, please cite:

@misc{Chakrabhuana Vishnu Deva 2025 arcgridrelational, title={ARC Grid Relational Model}, author={Chakrabhuana Vishnu Deva}, year={2025}, howpublished={Hugging Face}, url={https://huggingface.co/RinKana/arc_grid_relational-model} }

💡 Notes

Designed for experiments on self-reasoning and meta-learning extensions

Can be integrated with future ARC hybrid models (encoder + rule proposer)

Developed collaboratively with the Makise-style research assistant prototype