16 4 months ago

Tiny 125M AI Model

ollama run banaxitech/bananamind-v1

Details

4 months ago

0890bbf98dc4 · 271MB ·

gpt2
·
135M
·
F16
<|im_start|>system You are a helpful assistant.<|im_end|> <|im_start|>user {{ .Prompt }}<|im_end|> <
{ "stop": [ "<|im_end|>" ], "temperature": 0.8, "top_k": 50, "top_p": 0.

Readme

🍌 BananaMind V1

A small, local, and honestly not very good 125M parameter language model — but it runs fast, fits anywhere, and it’s yours.

⚠️ Fair warning: This is a personal project model. It will hallucinate, forget things after 512 tokens, and occasionally say something completely unhinged. You have been warned.


Quick Start

ollama run banaxitech/bananamind-v1

That’s it. No GPU required.


What Is This?

BananaMind V1 is a tiny conversational model fine-tuned on the OpenAssistant oasst1 dataset. It uses the ChatML prompt format and is designed for local, CPU-friendly inference.

Property Value
Parameters 125M
Context Window 512 tokens
Format GGUF
License Apache 2.0
Prompt Format ChatML

Modelfile

If you’re building on top of this model or tweaking settings, here’s the full Modelfile:

FROM ./BananaMind-V1.gguf

TEMPLATE """<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""

PARAMETER top_p 0.9
PARAMETER stop <|im_end|>
PARAMETER temperature 0.8
PARAMETER top_k 50

The newlines inside the TEMPLATE block must be real newlines, and stop must be set without quotes.


Usage Tips

  • Keep prompts short and direct — 512 tokens goes fast
  • Works best for simple Q&A and casual chat
  • Don’t trust it for facts, math, or anything important
  • Higher temperature = more creative (and more unhinged)

Limitations

  • Forgets context quickly (512 token window)
  • Makes things up with confidence
  • Struggles with multi-step reasoning
  • Not suitable for production use

License

Apache 2.0 — free to use, modify, and distribute.


Links


Small model. Big dreams. Low bar. 🍌