# Ares Seed 1M Model

A tiny real Ares checkpoint has been trained from scratch as a proof that the pipeline works end-to-end.

Model repo:

```text
https://huggingface.co/jacmor64/ares-seed-1m
```

Dataset repo:

```text
https://huggingface.co/datasets/jacmor64/ares-starter-dataset
```

## What this seed model is

- About 1.05M parameters.
- Decoder-only Transformer using the Ares code in this repo.
- BPE tokenizer trained from scratch.
- Continued to 5,000 CPU training steps.
- Best validation loss after continued training: about 4.42.
- Training data included a small streamed Simple Wikipedia sample, ML-process curriculum, and roleplay curriculum.

## What this seed model is not

It is not a capable assistant yet. It is intentionally tiny and undertrained. Its outputs are still rough and repetitive. Its purpose is to prove that Ares can be trained and generate from its own checkpoint weights without external AI APIs.

## Next step

Use `colab/Ares_Wikipedia_Roleplay_Brain_Training.ipynb` to train the stronger 30M Ares brain with more Wikipedia and roleplay data.