jacmor64/ares-seed-1m

Ares Seed 1M

This is a tiny proof-of-training checkpoint for Ares. It is not a capable chatbot. It proves the Ares pipeline can train a from-scratch decoder-only Transformer and generate from its own weights.

Training summary

• Architecture: Ares decoder-only Transformer
• Parameters: about 1.05M
• Context: 128 tokens
• Tokenizer: BPE trained from scratch, 2048 vocab
• Training: 5,000 CPU steps
• Data mixture:
  • Simple Wikipedia streamed from wikimedia/wikipedia
  • Ares machine-learning process curriculum
  • Ares roleplay curriculum
• Best validation loss after continued training: about 4.42

Files

• ares_seed_1m_inference.pt — slim PyTorch checkpoint containing model weights/config.
• tokenizer.json — trained BPE tokenizer.
• config.json — Ares model config.
• sample_generation.txt — sample outputs from the tiny seed model.
• training_report.html — offline training report.
• training_manifest.json — data mixture and training summary.

Use

Clone the Ares Static Lab Space repo and run:

python -m ares_core.generate \
  --checkpoint ares_seed_1m_inference.pt \
  --tokenizer tokenizer.json \
  --prompt "Roleplay as a machine-learning tutor" \
  --device cpu

Caveat

This checkpoint is intentionally tiny and undertrained. Its outputs are rough and repetitive. Use the Colab notebook in the Ares Static Lab to train the larger 30M+ brain on more Wikipedia and roleplay data.