README.md

metadata

title: README
emoji: 👁
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false

ShallowMind - Just like DeepMind, but way more stupid🧠

Hi there! My name is Alessandro, i'm a ai research engineer. ShallowMind is my workspace for training and experimenting with language models.
The name is playful, but the goal is straightforward: to build increasingly capable models while exploring new ideas in pretraining and reasoning.

---

Research Interests

• Information-theoretic pretraining
  Looking at ways to identify and prioritize the most informative tokens, to see whether current scaling laws can be adjusted. (Work in progress — I’ll share results once experiments are further along.)

• Reasoning models
  Testing approaches that improve step-by-step and compositional reasoning.

• Architectural variations
  Extending my training pipeline to support Mixture-of-Experts (MoE) and other non-standard components.

---

Current Work

• Built a custom pre-training pipeline and pre-trained a first model from scratch (~1B scale) as a proof of concept.
• Iterating on the pipeline to add MoE layers and information-gain–based logic.
• Next steps:
  • Fine-tune the first model into Promptasaurus-Zero.
  • Train Blahblahthron-7B as a larger-scale follow-up experiment.

---

Roadmap

• Share ablations and code from early experiments.
• Scale training to larger models.
• Document results on token selection and reasoning tasks.

---