README.md

---
library_name: python
license: mit
tags:
- hybrid-ai
- local-assistant
- python
- web-development
- math
- physics
- chemistry
- html
- css
- javascript
- aion
pipeline_tag: text-generation
---

# AION

![AION logo](assets/aion_logo.svg)

![AION architecture](assets/aion_architecture.svg)


AION is a tiny hybrid local assistant built in a constrained CPU environment. It unifies several learned and symbolic components into one entrypoint:

```python
from aion import generate
print(generate("hola"))
```

## What AION can do

- Chat greetings and basic conversation.
- Write Python snippets and functions.
- Create web pages/components with HTML, advanced CSS and vanilla JavaScript.
- Solve many math tasks:
  - arithmetic,
  - linear equations,
  - quadratics,
  - derivatives/integrals for simple polynomials,
  - statistics,
  - geometry,
  - trigonometry,
  - combinatorics,
  - interest,
  - unit conversion.
- Solve basic physics formulas:
  - F=ma,
  - kinetic/potential energy,
  - Ohm's law,
  - power,
  - density,
  - momentum,
  - wave speed.
- Basic chemistry:
  - common elements,
  - moles,
  - molarity,
  - ideal gas law,
  - pH.
- Basic biology/general knowledge:
  - photosynthesis,
  - cells,
  - DNA,
  - evolution,
  - algorithms,
  - databases,
  - internet,
  - machine learning.

## Download

You can download the complete ready-to-run package from the repository files:

```bash
git lfs install
git clone https://huggingface.co/VoidWalkercero/AION-1
cd AION-1
python aion.py "hola"
```

Or from Python:

```python
from huggingface_hub import snapshot_download
path = snapshot_download("VoidWalkercero/AION-1")
print(path)
```

A zipped copy is also included under `download/AION-1.zip`.

## Architecture

AION is not a transformer LLM. It is a merged hybrid model:

1. `neural_python_mind.py` — NumPy character-level GRU trained for Python syntax/style.
2. `real_python_learner.py` — character n-gram learned intent classifier + compositional Python generator.
3. `real_web_learner.py` — character n-gram learned web intent classifier + HTML/CSS/JS generator.
4. `unified_learning_ai.py` — unified router for chat, Python, web, math and science.
5. A small deterministic math/science solver layer.

## Usage

CLI:

```bash
python aion.py "create a responsive landing page with dark mode"
python aion.py "solve 2x + 5 = 17"
python aion.py "force mass 10 acceleration 2"
python aion.py "write code to keep numbers greater than 12"
```

Python:

```python
from aion import generate
print(generate("what can you do"))
```

## Evaluation

![AION benchmark snapshot](assets/aion_benchmark.svg)


Local evaluation results are in:

```text
results/aion_local_eval.json
results/aion_local_eval.md
```

Summary:

| Suite | Score |
|---|---:|
| chat sanity | 3/3 |
| Python generation sanity | 3/3 |
| Web generation sanity | 4/4 |
| Math/science sanity | 6/6 |
| GSM8K test sample 30 | 0/30 |

Important: these are **not official Hugging Face leaderboard results**. AION is not a standard `transformers` model and cannot be directly submitted to most official HF benchmark leaderboards without a custom evaluation adapter. The GSM8K sample result is included honestly and shows the current limitation on multi-step word problems.

For optional comparison with small HF models, see `benchmark/benchmark_compare_small_models.py` and `benchmark/SMALL_MODEL_COMPARISON.md`.

## Limitations

- Not a large language model.
- Not a Transformers `AutoModel` checkpoint.
- Strong on composed templates and formulaic tasks; weak on deep natural-language reasoning.
- GSM8K multi-step reasoning is currently poor.
- Web output is generated as inline HTML/CSS/JS snippets suitable for local preview, not production-audited code.

## Training data

AION uses generated local curricula plus downloaded GSM8K JSONL files from OpenAI's public grade-school-math repository when available:

```text
outputs/unified_learning_ai/online_datasets/gsm8k_train.jsonl
outputs/unified_learning_ai/online_datasets/gsm8k_test.jsonl
```