CoolFace
Modelpublic

ParallaxOpen/Vela-Lumen-15M

sourceHugging Facecc-by-nc-4.0updated 17d agoView on Hugging Face
0likes355downloads
Model Card

Vela 15M Preview

A 15M-parameter math content generator trained from scratch on a single 8GB laptop GPU.

Preview: This is an early version. More training data and better results coming soon.

Downloads

FormatFileSizeHow to Use
GGUF F16vela-lumen-15m-f16.gguf59 MBWorks with Ollama, llama.cpp, LM Studio
Safetensors FP16model-fp16.safetensors44 MBHalf precision, Python
Safetensors FP32model.safetensors89 MBFull precision, training

Quick Start (Ollama)

bash
wget https://huggingface.co/ParallaxOpen/Vela-Lumen-15M/resolve/main/vela-lumen-15m-f16.gguf
cat > Modelfile << EOF
FROM vela-lumen-15m-f16.gguf
TEMPLATE "{{ .System }}{{ .Prompt }}"
SYSTEM "You are a math tutor."
EOF
ollama create vela-15m-preview -f Modelfile
ollama run vela-15m-preview

What It Does

Math Content Generation:

  • —Generates math word problems and equations
  • —Generates homework-style questions
  • —Generates math explanations and tutorials
  • —Generates LaTeX formulas

Example:

Prompt: "Create a math problem about:"
Output: "The expression (x/m * x) - (x/m * x) = x.
        Recall that (x/m * x) - (x/m * x) = x.
        Therefore, we can solve for..."

ELO Rating

MetricValue
Estimated ELO~800-850
GSM8K0%
Params15.5M

Comparison

ModelParamsGSM8KELO
Claude Fable 5UnknownUnknown1507
Llama 3.2 1B1.26B7.6%1111
SmolLM2-135M135M1.4%N/A
Vela 15M Preview15.5M0%~800-850

Model Details

PropertyValue
DeveloperParallaxOpen
Parameters15,486,080
ArchitectureTransformer Decoder, GQA 4:2, SwiGLU, RMSNorm, RoPE
Context512 tokens
Vocab24,189
LicenseCC BY-NC 4.0

Training

  • —Pretrain: 7B tokens, 200K steps, loss 2.5, ~10h on RTX 5060
  • —SFT: 14K math samples, 5K steps, loss 3.32, ~30min on RTX 5060
  • —Total: ~10 hours on a single 8GB laptop GPU

What's Next

  • —More pretrain data (FineMath 50B tokens)
  • —Better SFT with distilled data
  • —Larger model (50M+ params)
  • —Improved math reasoning

Citation

bibtex
@software{vela-15m-preview,
  title={Vela 15M Preview: A 15M-param Math Content Generator},
  author={ParallaxOpen},
  year={2026},
  note={Preview version - trained on a single RTX 5060 Laptop GPU}
}