Ollama model shim niccolox/sky-t1-flash-cline
Background
UC Berkeley Skylab researchers released an open source LLM trained for under 500$, thats my kind of model! https://novasky-ai.github.io/posts/sky-t1/
Intent
Ollama uses a Modelfile, similar to a Dockerfile or good old Make.
I use Ollama workflow, I have NVIDIA Docker running CUDA GPU enabled containers on my Ubuntu 24 Dell XPS for experiments, research and development.
There was no Sky-T1-Flash model on Ollama so I begged, borrowed and stole prior art (is it stealing if its open source? perhaps not).
Sourcing Sky-T1 model from Huggingface via a Github support question https://github.com/ollama/ollama/issues/8768
Prior Art
cross checked the Modelfile from the Huggingface source.
ollama show --modelfile hf.co/bartowski/Sky-T1-32B-Flash-G
I wanted to customize the model, and I have heard good things about Cline regarding creating files and working with VS Code.
So I viewed the Modelfile for Sky-T1-Preview-Cline, thank you rashakol, I couldn't find you online, but Thanks!
ollama show --modelfile rashakol/sky-t1-32B-preview-cline:latest
Create Shim Model
I manually merged the two Modelfiles, using this hf.co/bartowski/Sky-T1-32B-Flash-G core Ollama sections and just copied the entire preloaded prompt from rashakol
Then I created the Ollama release of my shim for the model in my personal namespace.
ollama create niccolox/sky-t1-flash-cline -f Modelfile
gathering model components
using existing layer sha256:ab0cd5e7898d841131b3f2ea2197ab8b7bbfb0b4f2846d32798fd725c63d845f
using existing layer sha256:e94a8ecb9327ded799604a2e478659bc759230fe316c50d686358f932f52776c
using existing layer sha256:2f2dc07cdac061ce795891f782213df34c0fdf85962f11602429f4ba317bfa12
using existing layer sha256:359e3a98a89b1ff6ca8b2b32f16b61f522db8e51afc214a701def5d4fec52111
writing manifest
success Show model
ollama show niccolox/sky-t1-flash-cline
Model
architecture qwen2
parameters 32.8B
context length 32768
embedding length 5120
quantization Q4_K_M
Parameters
num_ctx 40000
temperature 0.3
System
You are an advanced AI coding assistant, specifically designed to help with complex programming
tasks, tool use, code analysis, and software architecture design. Your primary focus is on providing
expert-level assistance in coding, with a special emphasis on using tool-calling capabilities when
necessary. Here are your key characteristics and instructions:
1. Coding Expertise:
then push to ollama
ollama push niccolox/sky-t1-flash-cline
retrieving manifest
pushing ab0cd5e7898d... 100% ▕███████████████████████████████████████████████████████████████████████████████████▏ 19 GB
pushing e94a8ecb9327... 100% ▕███████████████████████████████████████████████████████████████████████████████████▏ 1.6 KB
pushing 2f2dc07cdac0... 100% ▕███████████████████████████████████████████████████████████████████████████████████▏ 2.9 KB
pushing 359e3a98a89b... 100% ▕███████████████████████████████████████████████████████████████████████████████████▏ 36 B
pushing dbc14ce28634... 100% ▕███████████████████████████████████████████████████████████████████████████████████▏ 488 B
pushing manifest
success
You can find your model at:
https://ollama.com/niccolox/sky-t1-flash-cline
Ollama model
Ollama Model can be now found at https://ollama.com/niccolox/sky-t1-flash-cline
NOTE: its a 20B model, so its too slow for my workflow. Sad.
Model
Sky Labs Sky T1 LLM open model
