metadata
license: mit
datasets:
- EdinburghNLP/xsum
base_model:
- mistralai/Mistral-7B-Instruct-v0.3
pipeline_tag: summarization
Lightweight QLoRA adapter trained on the EdinburghNLP/XSum dataset for abstractive news summarization.
The adapter fine-tunes 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().