Why the metadata is appended in the code data ?
#12
by
gagan001
- opened
The paper mentions that it appends the metadata of code files like
<reponame>REPONAME<filename>FILENAME<gh_stars>STARS\nCode<eos> with a probability of 0.2. How is this information helpful in training the model or how is it being utilized?
gagan001
changed discussion status to
closed
gagan001
changed discussion status to
open
For example you can use filename to indicate which language the code is going through the extension. You can also influence the generation content, we observed that adding <filename>solutions.py slightly improved the evaluation score on HumanEval. We also tried to see if conditioning on GitHub stars would make the model generate higher quality content but it didn't seem to impact the evaluation scores, that being said the impact/use cases for this metadata are still open for investigation.
loubnabnl
changed discussion status to
closed