Lightweight LoRA adapters trained on the EdinburghNLP/XSum dataset for abstractive news summarization.
The adapters fine-tune Mistral-7B-Instruct-v0.3 using ~50k training and 2k validation samples prepared via a summarization prompt.
Training used the following parameters:
r=8,
lora_alpha=16,
lora_dropout=0.1,
target_modules=["q_proj", "k_proj", "v_proj", "o_proj"]
Adapters can be merged into the base model for inference using PeftModel.merge_and_unload().
