4 ms·

That’s awesome! How’d you get such great code usage examples out of J? It almost seems like the code is properly related to the names. GAN code seems to look l

• thesephist 5y ago
  The code generated is most definitely related to the names/descriptions! To do this, I have to first generate the description then generate the code _from the descriptions_. The downside of this is that I can't parallelize text generation, but the upside is the code feels much more realistic. Here's the prompt I used (from that same file): The idea here was to give the model a prompt that felt like a tutorial or some kind, and try to minimize non-Python non-ML-y code. --- $MODEL_DESCRIPTION_FROM_EARLIER Let\'s use this model. The basic use case takes only a few lines of Python to run the inference. Here are the first few lines. ```python