MatBot 1.4

Overview

MatBot 1.4 is the successor to MatBot 1.3, built on the SmolLM2 360M base model and fine tuned for math reasoning. This release continues the exploration of how far small language models can be pushed on specialized mathematical benchmarks with limited training capacity.

Unlike MatBot 1.3, which collapsed into template repetition, MatBot 1.4 demonstrates meaningful improvement. While still a very small model with clear limitations, it achieves measurable accuracy on both GSM8K and MATH-500, showing better stability and inference behavior compared to its predecessor.

MatBot 1.4 remains an experimental research model, offering insight into training dynamics and the practical limits of small architectures.

Benchmark Comparison

Accuracy results based on sampled evaluation:

Although the absolute numbers are low, the difference is significant. MatBot 1.4 is the first model in the series to produce consistent and non-collapsed outputs on math tasks.

Goals

1. Test whether increased model capacity improves fine tuning outcomes on math reasoning tasks.
2. Explore whether narrow mathematical training can remain stable without collapsing into template loops.
3. Continue the development of the MatBot series as a practical study in training limits and small model behavior.

MatBot 1.4 is not intended for real world math problem solving or production level reasoning.

License

Released under the MIT License.

Summary

MatBot 1.4 is a modest but meaningful leap forward in the MatBot series. It moves beyond the collapse issues of its predecessor and demonstrates that small models can gain limited but interpretable mathematical ability when capacity is scaled appropriately. While still far from solving math reliably, MatBot 1.4 serves as a valuable research checkpoint on the path toward more capable compact models.