thau

THAU v2.0 - Self-Learning Language Model

THAU (Thinking, Helpful, Autonomous, Understanding) is a self-learning language model fine-tuned from TinyLlama-1.1B with specialized training in tool calling, reasoning, and Spanish.

Model Description

Capabilities

• Tool Calling: Native JSON-based function invocation
• Chain of Thought: Step-by-step reasoning for complex problems
• Image Generation: Prompt engineering for image generation
• Spanish Fluency: Natural and technical conversations
• Programming: Python, JavaScript, Java assistance

Training Data

Usage

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("luepow/thau")
tokenizer = AutoTokenizer.from_pretrained("luepow/thau")

# Chat format
prompt = """<|system|>
Eres THAU, un asistente AI inteligente y servicial.</s>
<|user|>
Hola, quien eres?</s>
<|assistant|>
"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With Ollama (Recommended)

ollama pull luepow/thau
ollama run luepow/thau

Tool Calling Format

THAU uses a JSON-based tool calling format:

<tool_call>{"name": "tool_name", "arguments": {"param": "value"}}</tool_call>

Available Tools

Example

User: What time is it?

THAU:

<tool_call>{"name": "get_current_time", "arguments": {}}</tool_call>

Limitations

• Model size limits complex multi-step reasoning
• May hallucinate on topics outside training data
• Tool calling accuracy varies by complexity
• Spanish is the primary language; English is secondary
• Best for simple to moderate complexity tasks

Training Details

• Full Training: 3,022 data points, 4,533 steps, loss 0.94
• Specialized v2.0: 297 examples, 745 steps, loss 0.43
• Hardware: Apple Silicon (MPS)
• Training Time: ~7 minutes for specialized phase

Citation

@misc{thau2024,
  title={THAU v2.0: A Self-Learning Language Model},
  author={Luis Perez (luepow)},
  year={2024},
  url={https://huggingface.co/luepow/thau}
}

Links

• Ollama: luepow/thau
• GitHub: luepow/thau

Acknowledgments

• Thomas & Aurora - Inspiration for the cognitive age progression system
• Claude (Anthropic) - AI pair programming partner
• TinyLlama Team - Excellent base model
• Hugging Face - Model hosting and transformers library

---

THAU v2.0 - Built with incremental learning and specialized training

Dedicated to Thomas & Aurora