Loading Model

Initializing...

Mini LLM JS

This LLM was trained by custom CUDA code without relying on any ML libraries (by Daniel Chermetz)

GPU Mode CPU Mode

Context:

Step 1: Load a Story

or enter a story number (0 - 377,000):

Enter Your Story Text

Note: Characters not in the vocabulary will be replaced with spaces.

Select Model Snapshot

Choose a model checkpoint to load:

Model 5 is the latest snapshot, but it's not always or even usually the best. Result quality varies by prompt—with earlier snapshots (1-4) often producing better or more interesting results for specific inputs.

Note: The most interesting aspect of this project isn't the result quality, but that these models were trained entirely using custom hand-written JavaScript and browser supported code—no PyTorch or other ML libraries were used. While the UI received AI assistance, the inference engine, BPE tokenization, and training were all hand-written.

Step 2: Choose Inference Mode

Select % of story context to provide:

Step 3: Generate

Teacher Forced Mode Modify Selection

Allow editing story context

Story Context:

Please load a story to begin...

Model Output:

Show token probabilities Show token rank

The model's response will appear here.

Select a New Story