Bring the evidence
Use reviewed text, record its sources, and separate training, validation and test files.
● AN INDEPENDENT LEARNING EXPERIMENT
Start at zero.
Learn how the
universe works.
Your own small language model, trained from random weights. Explore science, ask better questions, and keep evidence at the heart of every answer.
No pretrained model · No paid API · Portable checkpoints
● Notebook readyGPU: not connectedTraining: run in your notebookWeights: random initialization
01 / THE EXPERIMENT
A real beginning.
A modest first model.
Start with a tiny causal transformer. The science summaries test the pipeline. They are much too small to teach a useful language model.
Train from scratch
Every weight starts randomly. A byte vocabulary and causal attention learn to predict the next byte.
Question the result
Inspect held-out text and generated samples. A lower loss measures prediction, not truth.
Keep your progress
Export weights and checkpoints. Restore the exact data and settings to continue elsewhere.
PEPAMAN_UNIVERSE_FROM_SCRATCH.IPYNB
Your portable training workspace.
Import into Kaggle or Colab, choose a GPU, run the cells, and download your checkpoint before the session ends.
- Download the notebook
- Import / upload it in your provider
- Select an available GPU runtime
- Run cells in order; export your work
02 / KNOWLEDGE WITH CONTEXT
Evidence first.
Uncertainty included.
An observation, a scientific model, an open question and a philosophical belief deserve different labels.
Inspect your text
ON YOUR DEVICECheck a UTF-8 .txt file before adding it to the notebook. This checker reads your file locally in this browser; it never uploads it.
No file selected.
Starter corpus
DEMONSTRATIONOriginal, brief summaries with sources. Too small for a capable assistant. Review and expand before serious training.
TrainingCosmic history · Sun · black holes · answer style↓ValidationDark matter · for model selection↓Final testClimate · reserved for final evaluation↓Source notesAttribution, split descriptions and review date↓Sources support physical claims. Answer-style guidance is project-authored. No NASA affiliation.
03 / RESOURCEFUL COMPUTING
Use what is available.
Save what you learn.
Pepaman hosts the notebook. Computation happens in your chosen runtime, after you sign in. Opening this page does not connect a GPU.
Kaggle Notebooks
Free GPUs have weekly quotas, eligibility rules and availability limits. Your account's displayed quota is authoritative.
Read GPU limits ↗Google Colab
Free GPU access has dynamic limits. Sessions can end, and a particular GPU or uninterrupted runtime is never guaranteed.
Read runtime limits ↗Your own computer
Test a tiny configuration on CPU using the same code. Your own hardware avoids cloud quotas; hardware and electricity still cost money.
Download training code ↓No lifetime GPU promise. Free tiers can change. Use legitimate accounts within each provider's rules. Checkpoints allow sequential continuation; they do not combine providers into a single supercomputer. Provider information checked 19 October 2026.
THE GUIDING PRINCIPLE
“I don't know yet”
is a useful place to start.
A small experiment can teach us how learning works. A trustworthy science assistant takes much more data, testing and human review. Keep curiosity. Keep the evidence.