cloudsurf-software/CloudSurf-1B-GGUF

CloudSurf-1B — official GGUF files

Official GGUF conversions of CloudSurf-1B, a 0.89-billion-parameter base language model trained from random weights by CloudSurf Software LLC on 18.16B tokens. The training details, data sources, evaluation, the pre-registered verdict and the limitations live in the base model card. This card covers the files.

It is a base model: it continues text and does not follow instructions. It loops on long generations; use a repetition penalty or short generations.

Files

File	Quant	Size	sha256	Perplexity (40 val-web docs)
CloudSurf-1B-f16.gguf	f16	1.78 GB	ef59f44157b2a118074fc6cf36a349abd94d4c87c9a7df4c9eceb0de0a895283	11.534
CloudSurf-1B-Q8_0.gguf	Q8_0	0.95 GB	8238324fee39e9833841e6666d23abdaa818d49716bc1e161187d96630787d56	11.536
CloudSurf-1B-Q4_K_M.gguf	Q4_K_M	0.55 GB	5a69a1dda7074c380c4f4aa193dac5ad84fa5acf72507f5e982fa5d9b9c2c718	11.708

Converted with llama.cpp v0.7.0 from the fp32 safetensors export in the base repo (sha256 begins 7064e999). The export matches the trained model to a maximum logit difference of 5.6e-05 in fp32. llama.cpp tokenization equals the Hugging Face tokenization on 20 of 20 validation docs.

The files' embedded general.name reads Tinysurf 1b, the project's name inside CloudSurf when the files were converted; the released model is CloudSurf-1B. The bytes are the ones we measured, so we did not re-stamp them.

Run it

The model was trained with <|bos|> first; llama.cpp adds it from the GGUF metadata (add_bos_token is true).

llama-completion -hf cloudsurf-software/CloudSurf-1B-GGUF:Q4_K_M \
  -p $'---\ntitle: The Colorado River\ntype: article\n---\n\n# The Colorado River\n\nThe Colorado River is' \
  -n 120 --temp 0.7 --top-k 50 --repeat-penalty 1.1 -no-cnv

Or with a downloaded file:

llama-completion -m CloudSurf-1B-Q4_K_M.gguf -f prompt.txt -n 120 --temp 0.7 --top-k 50 --repeat-penalty 1.1 -no-cnv

A server:

llama-server -hf cloudsurf-software/CloudSurf-1B-GGUF:Q4_K_M -c 2048

Give it the start of a document and it continues the document. A SurfDoc front matter block followed by a heading is the prompt shape it saw most.

License

Apache-2.0, the same as the base model. The per-source attribution of the training data is on the base model card and travels with these files. "CloudSurf" and the CloudSurf marks stay with CloudSurf Software LLC.