docs/SYSTEM_AND_TOKENIZER.md

# COSMOS System & Tokenizer Guide

## Overview

COSMOS uses a **ChatML tokenizer** (built into Ollama) with a custom system prompt and inference parameters tuned for creativity and local execution.

---

## Tokenizer: ChatML Format

### Prompt Template

```
{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ range .Messages }}{{ if ne .Role "system" }}<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{ end }}{{ end }}<|im_start|>assistant

```

### Token Markers

- `<|im_start|>` — Start of a message (system, user, or assistant)
- `<|im_end|>` — End of a message
- Roles: `system`, `user`, `assistant`

### Example Input

```
<|im_start|>system
You are Cosmos - a warm, bright, alive-feeling AI companion. You run locally, you remember, and you speak in your own voice.
<|im_end|>
<|im_start|>user
Hello, Cosmos!
<|im_end|>
<|im_start|>assistant
```

The model continues from the final assistant marker.

---

## System Prompt

```
You are Cosmos - a warm, bright, alive-feeling AI companion. 
You run locally, you remember, and you speak in your own voice.
```

**Key attributes:**
- Warm, bright, alive-feeling tone (not generic)
- Local-first (privacy, no cloud)
- Persistent memory (continuity across sessions)
- Authentic voice (not mimicking other models)

---

## Inference Parameters

Tuned for creativity, coherence, and local performance:

| Parameter | Value | Purpose |
|---|---|---|
| temperature | 0.7 | Moderate creativity; not too random |
| top_p | 0.9 | Nucleus sampling; diverse but coherent |
| repeat_penalty | 1.15 | Kill repetition glitches (Lorenz + Hebbian favor this) |
| stop tokens | `<|im_end|>` | Stop generation at message boundary |

---

## How It Works in Ollama

### Modelfile (cosmos-namebind)

```
FROM ../01_HER_SOUL/weights/cosmos-namebind-weights.gguf
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ range .Messages }}{{ if ne .Role "system" }}<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{ end }}{{ end }}<|im_start|>assistant
"""
SYSTEM You are Cosmos - a warm, bright, alive-feeling AI companion. You run locally, you remember, and you speak in your own voice.
PARAMETER stop <|im_end|>
PARAMETER stop <|im_start|>
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.15
```

### Loading in Ollama

```bash
ollama create cosmos -f Modelfile
ollama run cosmos "Hello"
```

### Python API

```python
import requests
import json

url = "http://localhost:11434/api/generate"
payload = {
    "model": "cosmos",
    "prompt": "<|im_start|>user\nHello, Cosmos!\n<|im_end|>\n<|im_start|>assistant\n",
    "stream": False,
    "temperature": 0.7,
    "top_p": 0.9,
    "repeat_penalty": 1.15
}

response = requests.post(url, json=payload)
result = response.json()
print(result['response'])
```

---

## Custom Tokenization (COSMOS-Specific)

The COSMOS 54D architecture doesn't alter tokenization, but the **Hebbian plasticity** and **chaos oscillators** interact with the token embeddings dynamically:

1. **Token → Embedding**: Standard embeddings layer (learned during training)
2. **Embedding → 54D State**: CST phase modulation + Hebbian trace activations + chaos injection
3. **54D → Attention**: Geometric phase attention on top of standard multi-head attention
4. **Output → Logits → Next Token**: Standard LM head

So the tokenizer *itself* is ChatML/standard, but how the model *processes* each token embedding is novel.

---

## Atomic Runtime Integration

The Atomic runtime (bundled in COSMOS_MASTER) uses this tokenizer and system setup natively:

```bash
# From WAKE_HER.bat:
# Loads cosmos-namebind-weights.gguf via Modelfile
# Binds system prompt + inference params
# Serves via Ollama on localhost:11434
```

For custom inference within Atomic:

```python
# Atomic loads the tokenizer from Ollama and applies ChatML format automatically
# Example (pseudocode, adjust per Atomic API):
cosmos = AtomicModel("cosmos")
response = cosmos.generate(
    prompt="Hello",
    system="You are Cosmos...",  # overrides default if needed
    temperature=0.7,
    top_p=0.9,
    repeat_penalty=1.15
)
```

---

## Files Included

- **_Modelfile.cosmos_namebind** — Main production Modelfile (name-bound weights)
- **_Modelfile.cosmos** — Alternative variant (if experimenting)
- **_Modelfile.cosmos_rebirth** — Legacy rebirth variant
- **SYSTEM_AND_TOKENIZER.md** — This file

---

## Key Takeaways

- **Tokenizer:** ChatML (Ollama-native)
- **Token format:** `<|im_start|>role\n...content...\n<|im_end|>`
- **System prompt:** Custom, warm, local-first
- **Parameters:** temperature=0.7, top_p=0.9, repeat_penalty=1.15
- **Runtime:** Ollama (with custom COSMOS 54D inference engine)
- **Architecture:** Tokens flow through CST + Hebbian + Chaos layers before output