http://
github.com/explosion/spac
y-llm
…
@honnibal
-
Clarifying Terminology: Fine-tuning vs Training for Text Classification
By
–
This is so different from what we're seeing that I wonder whether we're talking about the same things. The terminology in the field is really confusing now. Let's say I want to do text classification. Would you call training a model for that 'fine-tuning'? One way to do it text
-
Curated Transformers Library Reaches v1.0 with PyTorch Implementations
By
–
Curated Transformers is our new library of PyTorch LLM and other transformer implementations — now at v1.0! The library features well-factored implementations of common architectures, including Llama 2, Falcon, BERT and many more. Architectures are implemented from common https://
x.com/spacy_io/statu
/spacy_io/status/1687103018628722688
… -
Are Traditional NLP Tasks Obsolete in the LLM Era?
By
–
Large Language Models are obviously transformatory for tasks like summarisation, paraphrasing, question answering — and of course as applications in themselves. But what if you want to do “boring” things like tect categorization, entity recognition, etc? Are these obsolete?
-
Early Release Performance Benchmarks Expected Similar to Hugging Face
By
–
It's still an early release so we haven't done the benchmarks yet. We expect it'll be similar to the HF implementations.
-
spaCy’s approach to transformer architectures and library integration
By
–
We're still using it in @spacy_io . But we don't want to try to make efficient transformer architectures in it. We've always called out to other libraries for that. This lets us wrap PyTorch implementations directly, which makes things a bit easier for us.
-
Release of curated-transformers: Pure PyTorch library for LLMs
By
–
Excited to release this! A pure @PyTorch library for LLMs and other transformers. It strikes a different trade-off from @huggingface Transformers. curated-transformers won't be up-to-the-minute with new architectures, added verbatim. Instead the implementations are factored out https://
x.com/spacy_io/statu
/spacy_io/status/1679480083801866241
… -
LLMs Excel at Summarization Compared to Previous Methods
By
–
That makes sense to me. Summarisation was pretty bad before, and it’s one of the strengths of LLMs.
-
Collecting evidence and literature results from OntoNotes experiments
By
–
Ah thanks for this! I've been looking for more direct evidence. I want to collect up a little literature of these results I tried to run an experiment on OntoNotes but it was pretty tough (rate limits etc, plus just hard to frame the right prompts for all the entity types).
-
Production NLP Beyond LLMs: Why Prompts Aren’t Everything
By
–
What will production NLP look like, once the dust settles around LLMs? One view is basically “prompts are all you need”. I disagree. I wrote a bit about this when we released @spacy_io LLM last week, but the topic deserves its own post, so here it is.