docs/source_en/BestPractices/Reranker.md

Reranker Training

SWIFT supports Reranker model training. Currently supported models include:

1. modernbert reranker model
   • ModelScope Hugging Face
2. qwen3-reranker model
   • 0.6B: ModelScope Hugging Face
   • 4B: ModelScope Hugging Face
   • 8B: ModelScope Hugging Face
3. qwen3-vl-reranker model
   • 2B: ModelScope Hugging Face
   • 8B: ModelScope Hugging Face

Implementation Methods

SWIFT currently supports two implementation methods for Reranker models, which have significant differences in architecture and loss function computation:

1. Classification Reranker

Applicable Models: modernbert reranker models (e.g., gte-reranker-modernbert-base)

Core Principles:

• Based on sequence classification architecture, adding a classification head on top of pre-trained models
• Input: query-document pairs, Output: single relevance score

2. Generative Reranker

Applicable Models: qwen3-reranker models (0.6B/4B/8B)

Core Principles:

• Based on generative language model architecture (CausalLM)
• Input: query-document pairs, Output: probability of specific tokens (e.g., "yes"/"no")
• Classification is performed by comparing logits of specific tokens at the final position

Loss Function Types

SWIFT supports multiple loss functions for training Reranker models:

Pointwise Loss Functions

Pointwise methods transform the ranking problem into a binary classification problem, processing each query-document pair independently:

• Core Idea: Binary classification for each query-document pair to determine document relevance to the query
• Loss Function: Binary cross-entropy
• Use Cases: Simple and efficient, suitable for large-scale data training

Environment variable configuration:

• GENERATIVE_RERANKER_POSITIVE_TOKEN: Positive token (default: "yes")
• GENERATIVE_RERANKER_NEGATIVE_TOKEN: Negative token (default: "no")

Listwise Loss Functions

Listwise methods transform the ranking problem into a multi-classification problem, selecting positive examples from multiple candidate documents:

• Core Idea: Multi-classification for each query's candidate document group (1 positive + n negative examples) to identify positive documents
• Loss Function: Multi-class cross-entropy
• Use Cases: Learning relative ranking relationships between documents, better aligned with the actual needs of information retrieval

Environment variable configuration:

• LISTWISE_RERANKER_TEMPERATURE: Softmax temperature parameter (default: 1.0)
• LISTWISE_RERANKER_MIN_GROUP_SIZE: Minimum group size, if the number of documents in the group is less than this value, the loss will not be calculated (default: 2)

Listwise vs Pointwise:

• Pointwise: Independent relevance judgment, simple training, but ignores relative relationships between documents
• Listwise: Learning relative ranking, better performance, more suitable for the essential needs of ranking tasks

The loss function source code can be found here.

Dataset Format

# LLM
{"messages": [{"role": "user", "content": "query"}], "positive_messages": [[{"role": "assistant", "content": "relevant_doc1"}],[{"role": "assistant", "content": "relevant_doc2"}]], "negative_messages": [[{"role": "assistant", "content": "irrelevant_doc1"}],[{"role": "assistant", "content": "irrelevant_doc2"}], ...]}
# MLLM
{"messages": [{"role": "user", "content": "<image>query"}], "images": ["/some/images.jpg"], "positive_messages": [[{"role": "assistant", "content": "<image>relevant_doc1"}]], "positive_images": [["/some/positive_images.jpg"]], "negative_messages": [[{"role": "assistant", "content": "<image><image>irrelevant_doc1"}], [{"role": "assistant", "content": "<image>irrelevant_doc2"}]], "negative_images": [["/some/negative_images1.jpg", "/some/negative_images2.jpg"], ["/some/negative_images3.jpg"]]}

Field Description:

• messages: Query text
• positive_messages: List of positive documents relevant to the query, supports multiple positive examples
• negative_messages: List of negative documents irrelevant to the query, supports multiple negative examples

Environment Variable Configuration:

• MAX_POSITIVE_SAMPLES: Maximum number of positive examples per query (default: 1)
• MAX_NEGATIVE_SAMPLES: Maximum number of negative examples per query (default: 7)

By default, MAX_POSITIVE_SAMPLES positive examples and MAX_NEGATIVE_SAMPLES negative examples will be extracted from each data item. Each positive example will be grouped with MAX_NEGATIVE_SAMPLES negative examples to form a group. Therefore, each data item will be expanded into MAX_POSITIVE_SAMPLESx(1 + MAX_NEGATIVE_SAMPLES) data points. If the number of positive/negative examples in the data is insufficient, all positive/negative examples will be used. If the number of positive and negative examples in the data exceeds MAX_POSITIVE_SAMPLES and MAX_NEGATIVE_SAMPLES, random sampling will be performed. IMPORTANT: The expanded data will be placed in the same batch. Therefore, the effective batch size on each device will be per_device_train_batch_size × MAX_POSITIVE_SAMPLES × (1 + MAX_NEGATIVE_SAMPLES). Please adjust your per_device_train_batch_size accordingly to avoid out-of-memory errors.

Training Scripts

Training scripts provided by ms-swift:

• Qwen3-Reranker/Qwen3-VL-Reranker
• Pointwise Classification Reranker
• Pointwise Generative Reranker
• Listwise Classification Reranker
• Listwise Generative Reranker

For inference scripts, please refer to here.

Advanced

• Qwen3-Reranker Custom Instruction:
  • Default template:

<|im_start|>system
Judge 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|>
<|im_start|>user
<Instruct>: {Instruction}
<Query>: {Query}
<Document>: {Document}<|im_end|>
<|im_start|>assistant
<think>

</think>

• Default instruction:

  • Given a web search query, retrieve relevant passages that answer the query

• Instruction priority (nearest wins):

  • system inside positive_messages/negative_messages > system in main messages > default instruction.
  • That is, if a positive/negative message sequence contains a system, it takes precedence; otherwise, if main messages has a system, use it; if neither is provided, use the default instruction.