README.md

---
language:
  - en
license: mit
library_name: julia
tags:
  - julia
  - character-level
  - philosophy
  - scalar-autograd
  - pure-julia
  - scriptio-continua
  - text-generation
pipeline_tag: text-generation
datasets:
  - LisaMegaWatts/juliagpt-data
model-index:
  - name: JuliaGPT
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          type: LisaMegaWatts/juliagpt-data
          name: juliagpt-data
        metrics:
          - type: loss
            value: 2.34
            name: Val Loss
            verified: false
---

# JuliaGPT

An experimental **8,096 parameter** character-level GPT in pure Julia with scalar autograd. Explores minimal vocabularies inspired by ancient Greek *scriptio continua*. No external ML framework dependencies.

## Model Lineage

| Model | Params | Vocab | Context | Val Loss | Notes |
|-------|--------|-------|---------|----------|-------|
| [MicroJulia](https://huggingface.co/LisaMegaWatts/MicroJulia) | 4,992 | 27 chars | 64 | 2.43 | First proof-of-concept |
| **JuliaGPT** | **8,096** | **29 chars** | **256** | **2.34** | **Expanded context + vocab** |
| [JuliaGPT-v2](https://huggingface.co/LisaMegaWatts/JuliaGPT-v2) | ~10M | 38 chars | 256 | 2.91 | Scaled-up char-level |

## Architecture

| Parameter | Value |
|-----------|-------|
| Architecture | 1-layer Transformer (pure Julia, scalar autograd) |
| Parameters | 8,096 |
| Embedding dim | 16 |
| Layers | 1 |
| Attention heads | 4 |
| Head dim | 4 |
| FFN hidden dim | 64 |
| Context length | 256 characters |
| Vocabulary | 29 characters (a-z, space, period, + BOS) |

### Vocabulary

29 tokens: `` .abcdefghijklmnopqrstuvwxyz`` + BOS

Numerals converted to words, all punctuation removed except period.

## Training

| | Value |
|---|---|
| Dataset | Aristotle's Rhetoric + Euclid's Elements (8,461 chunks) |
| Best val loss | 2.34 |
| Framework | Pure Julia (scalar autograd, no Flux/Lux) |

## Files

| File | Description |
|------|-------------|
| `best_model.json` | Original model weights + optimizer state (JSON format, scalar autograd) |
| `vocab.json` | 38-character vocabulary array |
| `data/aristotle_rhetoric.txt` | Training data |

**Note:** The `.jld2` checkpoint files in this repo contain a different, larger model (384d/6L/38vocab). That model has been moved to [JuliaGPT-v2](https://huggingface.co/LisaMegaWatts/JuliaGPT-v2). The original JuliaGPT is preserved in `best_model.json`.

## Inference Settings

| Parameter | Value |
|-----------|-------|
| vocab_size | 29 |
| context_length | 256 |

## Provenance

- **Author**: LisaMegaWatts
- **Source code**: [DavinciDreams/JuliaGPT](https://github.com/DavinciDreams/JuliaGPT)
- **Training data**: [LisaMegaWatts/juliagpt-data](https://huggingface.co/datasets/LisaMegaWatts/juliagpt-data)

## License

MIT