Hi, thanks for your awesome work.
Reading through the paper, it says for the CI build repair task:
To prompt the models to solve the task, we use the following strategy. To prepare an
instruction, we locate the first occurrence of the case-insensitive substring “error” in
the logs and take a 7-line context around this occurrence (3 lines before and after).
If the substring is not found, we use 7 last log lines. The instruction then reads as
follows: “Fix CI in order for tests to pass. Relevant logs: {relevant_logs}”. We then
prompt the LLM to modify the code blocks provided by an oracle to align with the
given instructions, and pass all the code blocks in a single request in the following
way:
[start of file.py#L12]
...code line 12...
...code line 13...
...
[end of file.py#L12]
But I can't seem to find exact code that produces prompts formatted this way for an LLM in the provided benchmark code files. Could you please point to the place in the code where prompts are created?
Hi, thanks for your awesome work.
Reading through the paper, it says for the CI build repair task:
But I can't seem to find exact code that produces prompts formatted this way for an LLM in the provided benchmark code files. Could you please point to the place in the code where prompts are created?