The flakiness is fundamental, which is one reason supervised models will be better for production where possible imo. I think there needs to be good testing in place and potentially resampling the LLM if there's a parse failure. We're planning stuff to make that easier
@honnibal
-
spaCy LLM Extension: Integrate Large Language Models into spaCy Pipelines
By
–
Today we've published spaCy LLM, an extension to @spacy_io that lets you use LLMs to power spaCy components. If you've already been using LLMs in this sort of pipeline-like way, you should definitely give it a go. We have lots more in the works!
-
LLMs for Annotation, Distillation, Interpretability and Production ML
By
–
You can use the LLM for annotation or distillation, to help you train the components. LLMs should also be very helpful for interpretability and monitoring, and other tricky parts of production ML that make up much of the total project costs.
-
LLM Limitations: Knowledge Gaps and Performance Constraints
By
–
LLMs obviously feature extremely advanced natural language understanding, but for many many applications, you can't just ask the LLM your question and get out the answer. The knowledge will often be missing, or the scale and latency requirements would be infeasible.
-
Using LLMs for Pipeline Development Without Training Data
By
–
But even if you can't use LLMs at runtime for your problem, they can be extremely helpful during development. You can prototype a whole pipeline with no training data, letting you sketch everything out before the components are in place.
-
Common ML Applications: Recommendations, Fraud Detection, Content Tagging
By
–
Other common applications are now completely pervasive. Things like recommendation systems, fraud or abuse detection, content tag suggestions and more are all still really difficult to do well, but they're definitely not uncommon.
-
Using LLMs for Traditional NLP Business Problems Effectively
By
–
A lot of people are building truly new things with LLMs, like wild interactive fiction experiences that weren't possible before. But if you're working on the same sort of NLP problems that businesses have been trying to solve for a long time, what's the best way to use LLMs? https://
x.com/spacy_io/statu
/spacy_io/status/1656734286425255937
… -
Loading spaCy models from directories and Hugging Face Hub
By
–
You can also pass a directory to spacy.load() if that's more convenient to you. It's also available on the HF hub.
-
Aspect-based Sentiment Analysis: Modeling Approaches and Techniques
By
–
Aspect-based sentiment is stuff like “service: good, food: bad”. Sometimes you can model it as multiple label vectors on the text, other times it’s more like spancat or relation extraction
-
Terminology Lists for NER and Information Extraction Tasks
By
–
Collecting terminology lists is a very useful step for most NER or information extraction tasks. It lets you annotate more quickly and consistently. It also helps you compute useful baselines, by evaluating how a rule-based approach would perform on the task.