Repo-level FIM format

[kevmo314]

Hi there, the technical report for Qwen2.5-Coder indicates that the repo-level FIM format follows:

<|repo_name|>{repo_name}
 <|file_sep|>{file_path1}
 {file_content1}
 <|file_sep|>{file_path2}
 {file_content2}
 <|file_sep|>{file_path3}
 <|fim_prefix|>{code_pre}<|fim_suffix|>{code_suf}<|fim_middle|>{code_fim}<|endoftext|>

However, the technical report also references Lozhkov, et al (2024) which follows a slightly different format:

<repo_name>reponame<file_sep>filepath0\ncode0<file_sep><fim_prefix>filepath1\n
 code1_pre<fim_suffix>code1_suf<fim_middle>code1_mid<file_sep> ...<|endoftext|>

Specifically, Lozhkov sets the FIM prefix token before the filepath to be completed, whereas the technical report seems to suggest Qwen2.5 was trained with the token after. I want to confirm that this is correct, that the repo-level FIM indeed follows a slightly different format?

[jklj077]

cc: @huybery