squishy

Squishy 150m

<p align="center"> <img src="squishy.jpg" alt="Squishy" width="400"/> <br> <em>Squishy — 109M parameter transformer, trained from scratch on consumer hardware</em> </p>

Created/Trained from "scratch" locally. Its framework sets up the architecture for future models and specialists. This is a COMPLETIONIST model. That means, you start a sentence, it will finish it. While it works ok for single turns + "Once upon a time", it can go off track with multiple-turn "conversations." Baby brain. Baby memory and this one is not chat tuned.

Temp does matter, or you will get a "Ow, bad Squishy!" <br><br>

0.8 Temp - Little Lilly loved a mouse and then buried it alive. <br>

Once upon a time, there was a little girl named Lily. She loved to play outside and explore the world around her. One day, she found a small hole in the ground. She looked inside and saw that there was a little mouse inside. Lily wanted to help the mouse, so she gently put it in her pocket and took it home. She made a little house for the mouse and gave it some food and water. The mouse was very happy and grateful. After a few days, the mouse's family was very busy. They had to work hard all day long. Lily wanted to help, but she was too little. So, she decided to bury the mouse in the ground. She dug a hole in the ground and carefully put the mouse inside. She covered the hole with dirt and patted it down. Lily felt happy that she could help the mouse and make the world a little better.

0.7 Temp - Little Lilly actually made a coherent (very) childlike response. <br>

Once upon a time, there was a little girl named Lily. She loved to play outside in her backyard. One day, she saw a snake slithering around in the grass. She was scared and ran to her mommy. "Mommy, mommy! There's a snake in the grass!" Lily said. Her mommy came outside and saw the snake too. "Don't worry, Lily. The snake is harmless. It won't hurt you," her mommy said. Lily felt better and went back to playing. Later that day, Lily saw her friend Billy in her backyard. "Hi Billy!" Lily said. "Do you want to play with me?" Billy said yes and they played together. They had so much fun that Lily forgot about the snake. She was happy that she was brave enough to play with her friend.

<p align="center"> <img src="ollama-example-chat.jpg" alt="Squishy" width="400"/> </p> <p align="center"> Prompt: Once upon a time </p> <p align="center"> <img src="deltakitsune.jpg" alt="delta:kitsune:forge" width="600"/> </p>

Kitsune TinyStories 150M

A small decoder-only transformer trained from scratch in the Kitsune Fine Tuning Suite as a proof-of-concept consumer-hardware pipeline run. This model was trained on TinyStories-style data to validate the end-to-end workflow: tokenizer training, raw PyTorch training loop, checkpointing, sampling, Hugging Face export, GGUF export, and Ollama deployment.

This is a small completion model, not a general-purpose assistant and not a chat-tuned model. It works best for short story generation, simple continuations, and lightweight creative experiments. Output quality is sensitive to sampling temperature; around 0.7 is a good default, while higher values may become unstable or surreal.

Recommended usage:

• short story prompts
• completion/generation tasks
• educational or pipeline demonstration use
• small-model experimentation

Not intended for:

• factual QA
• long-form reasoning
• reliable instruction following
• safety-critical use

Architecture:

• decoder-only transformer
• approximately 150M parameters
• modern small-model design with RoPE, RMSNorm, and SwiGLU
• trained locally on consumer GPU hardware

Suggested inference settings:

• temperature: 0.7
• top_p: 0.95
• keep prompts short and concrete

Lineage:

• created and trained in the Kitsune Fine Tuning Suite
• exported as a Hugging Face-compatible model and GGUF for local deployment