LeLM-GGUF

GGUF quantization of KenWu/LeLM, an NBA take analysis model fine-tuned on Qwen3-8B.

Available Quantizations

File	Quant	Size	Description
LeLM-Q4_K_M.gguf	Q4_K_M	4.7 GB	Best balance of quality and size

Usage with Ollama

Create a Modelfile:

FROM ./LeLM-Q4_K_M.gguf

PARAMETER temperature 0.7
PARAMETER top_p 0.9

SYSTEM You are LeLM, an expert NBA analyst. Fact-check basketball takes using real statistics. Be direct, witty, and back everything with numbers.

Then run:

ollama create lelm -f Modelfile
ollama run lelm "Fact check: LeBron is washed"

Usage with llama.cpp

llama-cli -m LeLM-Q4_K_M.gguf -p "Fact check this NBA take: Steph Curry is the GOAT" -n 512

Model Details

• Base model: Qwen3-8B
• Fine-tuning: LoRA (r=64, alpha=128) with SFT on NBA take analysis data
• Training: 3 epochs, 915 steps, final loss 0.288
• LoRA adapter: KenWu/LeLM

Part of LeGM-Lab

This model powers LeGM-Lab, an LLM-powered NBA take analysis and roasting bot.