RAW job posting data: https://huggingface.co/datasets/xanderios/linkedin-job-postings Run LLM extraction on the raw data for skills Validate this vs the skillspan dataset That gives you precision/recall numbers on your extraction prompt before you've trained anything. This is your LLM baseline.
https://github.com/kris927b/SkillSpan/
Use pretained NER model https://huggingface.co/docs/transformers/en/model_doc/distilbert Fine-tune it on the LLM generated training data, evaluate it on the skill span dataset
Evaluate the difference in accuracy vs cost trade off