metadata
base_model: HuggingFaceTB/SmolLM-135M
datasets:
- LDJnr/Capybara
###EVEN SMALLER Frankenstein of smolLm-0.13b upped to 0.15b Use this frankenbase for training.
If you're here from twitter and imatient, get the trained checkpoint file.
biggie-smollm-checkpoint-twitter-q8_0.gguf
wget https://huggingface.co/nisten/Biggie-SmoLlm-0.15B-Base/resolve/main/biggie-smollm-checkpoint-twitter-q8_0.gguf
./llama-cli -n 1024 -fa -b 512 --min-p 0.3 --top-p 0.85 -ctk q8_0 -ctv q8_0 --keep -1 -p "You are a Nasa jpl engineer teaching the user about space and cats. <|im_start|>User: How to build a city on Mars via calculating Aldrin-Cycler orbits?<im_end> /n " -m biggie-smollm-checkpoint-twitter-q8_0.gguf --temp 2 -ngl 0 -t 1 -co -cnv --reverse-prompt "Assistant:"
Done via semi-automated continuous merging to figure out the recipe. Model is more coherent.
wget https://huggingface.co/nisten/Biggie-SmoLlm-0.15B-Base/resolve/main/Biggie_SmolLM_0.15B_Base_bf16.gguf
llama-cli -ngl 99 -co --temp 0 -p "How to build a city on Mars via calculating Aldrin-Cycler orbits?" -m Biggie_SmolLM_0.15B
_Base_bf16.gguf
The temperature settings and min p etc need to be adjusted but even at default temp0 it was coherent for first 100 tokens. Amazing option for further training. And this is a merge of the base, not the instruct!
I don't understand how the f a 150mb file can talk but it can