SmolLM2-135M Prompt Enhancer (private)

A LoRA fine-tune of SmolLM2-135M-Instruct that expands short image-generation prompts into detailed descriptions. The adapter is merged into the weights.

Usage

System prompt:

User message format (required):

Files

• Root: merged safetensors model and tokenizer (Transformers)
• gguf/smollm2-prompt-enhancer-q8_0.gguf: quantized model for llama.cpp, Ollama, LM Studio
• gguf/Modelfile: Ollama configuration

Training

• Dataset: gokaygokay/prompt-enhancer-dataset (16,110 train / 500 eval sampled from 1,790 test)
• LoRA r=8, alpha=32, targets q/k/v/o projections
• 3 epochs, effective batch 16, lr 2e-4 cosine, bf16 on NVIDIA L4