Lack of Instruction & Different System Prompt in Qwen3 Reranker

https://github.com/StarlightSearch/EmbedAnything/blob/95f56eb3a07dca63a982f1df4dfd76babc1766b1/rust/src/reranker/model.rs#L158 According to the page: https://huggingface.co/Qwen/Qwen3-Reranker-0.6B Our approach: ```rust let prefix = "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be yes or no.<|im_end|>\n<|im_start|>user\n"; let suffix = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"; ``` is slightly different from the reference implementation: ```python3 prefix = "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n<|im_start|>user\n" suffix = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n" ``` the `yes` and `no` are not quoted, which may cause slightly lower probability scores. Given that the reranker is trained with exactly the same system prompt, it's best to keep it perfectly aligned with the official approach. Also, there's a larger issue: ```rust let pairs = queries .iter() .flat_map(|query| documents.iter().map(move |doc| (*query, *doc))) .collect::<Vec<_>>(); let pairs = if is_qwen3 { pairs.iter().map(|(query, doc)| (format!("{}{}", prefix, query), format!("{}{}", doc, suffix))).collect::<Vec<_>>() } else { pairs.iter().map(|(query, doc)| (query.to_string(), doc.to_string())).collect::<Vec<_>>() }; ``` And correct me if I'm wrong, the pair then get sent to be tokenized straightaway, without further formatting. However, in the official reference implementation: ```python3 def format_instruction(instruction, query, doc): if instruction is None: instruction = 'Given a web search query, retrieve relevant passages that answer the query' output = "<Instruct>: {instruction}\n<Query>: {query}\n<Document>: {doc}".format(instruction=instruction,query=query, doc=doc) return output ``` If it's not formatted in the way how it was trained, the LLM can only try to split the query and document based on its semantic understanding, since there's no structural format to split the query and doc. A workaround is to do the formatting before passing into EmbedAnything (by combining instruction & query as a new query, and also formatting the doc), but I would really love to see if there could be an official support.