Uri Alon

Hi @Zadagu , Thank you for your interest in SLM. I understand the confusion - the order of targets is not their sequential order. Specifically, it seems that `idleSources` can...

Hi @Zadagu , Sure, you can ask as many questions as you like :-) Let me separate my answer into two parts: ## Training If `idle,sources` appears in the context,...

Hi @Zadagu , I agree that this is not clear from the paper. The dimensions of `\tilde{h}`, `E^{type}` and `E^{subtoken}` are all equal (they must be, as you noticed) and...

Hi @Zadagu , Sorry for the late reply. I'm not sure why I did not mention in the paper, but I think that the source of discrepancy is as follows:...

Hi Zadagu, Thank you for reporting this. Yes, I am using the projected paths everywhere, including the copy mechanism, not only in the transformer. But I found another possible difference:...

Hi Zadagu, >Ci with a size of 9 x 256 x 256 In my implementation, the first dimension is `8` instead of `9` (that is: `8 x 512 x 512`)....

Hi @liu1234567yi , Thank you for your interest in our work! We have [this script](https://github.com/tech-srl/slm-code-generation/blob/master/scripts/collect_grammar.py) that as a preprocessing step, goes through the training data and finds all the terminal...

Hi Son, Thank you for your interest in our work! We eventually decided to release only the trained-model demo + API. We also released the data and the data creation...

Sure! Feel free to ask as many questions as you'd like. Sorry for the missing file, I just uploaded it. Let me know if you have any more questions.

Hi @ocramz , Sorry about that. Let me know if there are any specific things that are unclear. We have a newer model, PolyCoder, that performs better, and works for...