Natural Language Processing

Setup

Currently a combination of regular expressions and LLMs.

Regex Setup

Most regular expressions are straightforward and can be executed as Python code in the Features table described below.

LLM Setup

Download InstructionsForDownloadingLLaMALLMmodel.docx

Under the nlp/llm_code folder:

• This contains a datasets folder with scraped datasets, in addition to some labeled datasets loaded for the fine-tuned model. The output sample datasets are also contained here. Notably, CS Vape and Vape WH are available as sample outputs.

• Most functions can be executed through nlp/llm_code/llama_vape_csvape_test.ipynb.

  • Some NLP functions are successfully loaded through regex functions in the regex_functions.py file. These include e-liquid contents and nicotine levels.

  • Other functions that use the LLM code, including prompts, are available in llm_functions.py. Currently, the LLM code uses a fine-tuned version of Meta’s Llama-3.1-8B-Instruct model. Instructions for setting up this model are found under doc/Instructions for downloading LLaMA LLM model.docx.

  • An output viewer for a better understanding of the NLP output is available in output_explorer.ipynb.

NLP Features

Feature Name	Description	Code Location	Sample Data	Notes
Flavors	Currently, flavors is captured using regex and completed for common patterns in vape.com and vapewh. We are in the process of implementing LLM prompting to extract the flavors. Flavors currently are stored in a dictionary data structure, with the key being the flavor name and value being the description.	nlp/llm_code/regex_functions/extract_flavors_with_descriptions, nlp/flavor (MyVaporStore)	nlp/llm_code/datasets/output/processed_output, nlp/flavor/myvaporstore_flavors.csv	Since data is not consistent across different sources, we are working to standardize it. LLM will assist in standardizing the data for easier parsing and storage.
Screens	A regular-expression-based script to detect various screen features: display_type, color_display, touch_screen, curved_screen, battery_indicator, eliquid_indicator, smart_display, digital_display, hd_display, animated, backlit.	nlp/screens.py	nlp/screens_sample_data	Does not capture all aspects of “gaming” features, which will be part of another script.
Product Type	Product type is captured using LLaMA-based classification. Categories include Closed Refills, Closed System, Disposable System, E-liquid, and Accessories. Further information can be found in doc/Vape_Product_Categories.docx.	nlp/llm_code/llm_functions/classify_product	nlp/llm_code/datasets/output/processed_output	Requires consistent labeling of categories and may need LLM fine-tuning for specific outliers or new product types. csvape and vapewh have labeled datasets for reference (nlp/llm_code/datasets/labeled).
Iced/Menthol	This is in progress–we will continue work on this following completion of flavor parsing. This may be completed either using regex or LLaMA-based classification pending additional investigation.	TBD	TBD
Total Ounces/mL	Captured via regex to extract volume values (e.g., ounces or mL) from product descriptions.	nlp/llm_code/regex_functions/find_eliquid_contents	nlp/llm_code/datasets/output/processed_output	Multiple volumes may be available for some products. Additional work can be done to handle this similar to nicotine levels.
Nicotine Level	Captured via regex patterns to extract nicotine levels (e.g., 0mg, 3mg). Multiple levels are stored in separate columns.	nlp/llm_code/regex_functions/find_nicotine_levels	nlp/llm_code/datasets/output/processed_output
Synthetic Nicotine	Synthetic nicotine is detected using LLaMA-based classification to identify key terms (e.g., “tobacco-free nicotine”).	nlp/llm_code/llm_functions/classify_tfn	nlp/llm_code/datasets/output/processed_output	LLM captures most of the edge cases–may need additional prompting if any new verbiage is found. csvape and vapewh have labeled datasets for reference (nlp/llm_code/datasets/labeled).
Nicotine Free	Uses Nicotine Level to indicate if the product is nicotine free or not alongside relevant verbiage.	nlp/llm_code/regex_functions/find_nic_free	nlp/llm_code/datasets/output/processed_output	Additional edge cases may warrant LLM use. csvape and vapewh have labeled datasets for reference (nlp/llm_code/datasets/labeled).
CBD/THC	CBD/THC is detected using LLaMA-based classification. Zero-shot learning (no examples or additional training) has been successful in classifying CBD for products available.	nlp/llm_code/llm_functions/classify_cbd	nlp/llm_code/datasets/output/processed_output	Larger test dataset may be useful to obtain a more robust accuracy metric. csvape and vapewh have labeled datasets for reference (nlp/llm_code/datasets/labeled).