Project Overview

Experiments on a compact artificial token language (ATL) trained with SentencePiece and evaluated against GPT-style BPE for efficiency and reasoning.

Key Findings

• ATL-512 increased token density by ~112% vs. cl100k_base on WikiText-103 (0.543 vs. 0.256 tokens/char).
• ATL-coded GSM8K prompts were ~3× longer than English (274 vs. 92 tokens on average).
• Using a small local model (distilgpt2), reasoning accuracy was 0/12 for both English and ATL; no evidence of quality preservation.
• Conclusion: this ATL configuration hurts efficiency; better coding and stronger models are needed.

How to Reproduce

1. Ensure Python 3.12+ and uv are available.
2. From repo root: uv venv && source .venv/bin/activate
3. Install deps: uv sync (or pip install -r requirements.txt).
4. Run experiments: python notebooks/atl_experiments.py
5. Outputs: metrics in results/, plots in results/plots/.

File Structure

• planning.md — research plan.
• notebooks/atl_experiments.py — tokenizer training, efficiency stats, reasoning eval.
• results/ — CSV/JSON metrics and plots.
• artifacts/atl_512.model|vocab — trained ATL tokenizer.
• REPORT.md — full report with methods and findings.

See REPORT.md for detailed analysis and discussion.