ms-swift/docs/source_en/BestPractices/Embedding.md

# Embedding Training

SWIFT has already supported the training of embedding models, including both pure text and multimodal types. Currently supported models include:

1. modernbert embedding model
   - [ModelScope](https://modelscope.cn/models/iic/gte-modernbert-base) [Hugging Face](https://huggingface.co/Alibaba-NLP/gte-modernbert-base)
2. gte embedding models
   - 1.5B: [ModelScope](https://www.modelscope.cn/models/iic/gte_Qwen2-1.5B-instruct) [Hugging Face](https://huggingface.co/Alibaba-NLP/gte-Qwen2-1.5B-instruct)
   - 7B: [ModelScope](https://www.modelscope.cn/models/iic/gte_Qwen2-7B-instruct) [Hugging Face](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct)
3. gme embedding models
   - 2B: [ModelScope](https://www.modelscope.cn/models/iic/gme-Qwen2-VL-2B-Instruct) [Hugging Face](https://huggingface.co/Alibaba-NLP/gme-Qwen2-VL-2B-Instruct)
   - 7B: [ModelScope](https://www.modelscope.cn/models/iic/gme-Qwen2-VL-7B-Instruct) [Hugging Face](https://huggingface.co/Alibaba-NLP/gme-Qwen2-VL-7B-Instruct)
4. qwen3-embedding models
   - 0.6B: [ModelScope](https://www.modelscope.cn/models/Qwen/Qwen3-Embedding-0.6B) [Hugging Face](https://huggingface.co/Qwen/Qwen3-Embedding-0.6B)
   - 4B: [ModelScope](https://www.modelscope.cn/models/Qwen/Qwen3-Embedding-4B) [Hugging Face](https://huggingface.co/Qwen/Qwen3-Embedding-4B)
   - 8B: [ModelScope](https://www.modelscope.cn/models/Qwen/Qwen3-Embedding-8B) [Hugging Face](https://huggingface.co/Qwen/Qwen3-Embedding-8B)

Developers can integrate their own models by ensuring the model forward output satisfies:

```text
{"last_hidden_state": some-embedding-tensor}
```

The return value should be a JSON with a `last_hidden_state` key, where the value is an embedding tensor. For the input part, you can use our already supported templates. Users can also specify the

```shell
   --task_type embedding
```
parameter to convert any other model into an embedding model for training.

It should be noted that the embedding models currently supported by SWIFT are all based on pure text or multimodal LLMs, and CLIP-type model training is not currently supported.

Additionally, all embedding models supported by SWIFT have normalization added at the end of the model forward pass. If you add new models yourself, please remember to include a normalization layer.

## Loss

The Embedding models supported by SWIFT currently can use the following loss functions:

- **cosine_similarity**: Cosine similarity loss, which calculates the similarity between two embeddings and fits based on the label value. It is effectively an MSE loss.
- **contrastive**: Contrastive learning loss with adjustable margin. Labels are only supported as 0 and 1.
- **online_contrastive**: Contrastive loss considering hard negatives and hard positives. Labels are only supported as 0 and 1.
- **infonce**: Computes pairwise cosine similarities between different rows within the same batch, maximizing similarity within rows and minimizing similarity between different rows. No labels are required.

The source code for the loss functions can be found [here](https://github.com/modelscope/ms-swift/blob/main/swift/plugin/loss.py).

## Dataset Format

> **Note:**
> 1. The `<image>` tag in the multimodal section below can appear in any position within `query`, `response`, or `rejected_response`. It is only required that the number of tags matches the number of values in `images`.
> 2. The correspondence between tags and `images` follows the order: first matching the `<image>` tags in `query`, then those in `response`, and finally parsing the `<image>` tags in `rejected_response` sequentially.
> 3. `query` represents the anchor sample, `response` represents the positive or contrastive sample, and `rejected_response` corresponds to hard negative samples.
> 4. The `<video>` and `<audio>` tags are also supported, enabling native support for video and audio embeddings.

### Format for Cosine Similarity Loss

```json lines
# LLM
{"query": "sentence1", "response": "sentence2", "label": 0.8}
# MLLM
{"query": "<image>", "response": "<image>sentence", "images": ["/some/images1.jpg", "/some/images2.jpg"], "label": 0.7}
{"query": "sentence1", "response": "<image>sentence2", "images": ["/some/images1.jpg"], "label": 0.7}
```

The eval metrics are the Pearson and Spearman's Rank Correlation Coefficient of the embeddings' euclidean distance/dot production and so on, totally 8 values.

### Format for Contrastive/Online Contrastive Loss

```json lines
# LLM
{"query": "sentence1", "response": "sentence2", "label": 1}
# MLLM
{"query": "<image>", "response": "sentence", "images": "/some/images.jpg", "label": 1}
{"query": "<image>sentence1", "response": "sentence2", "images": "/some/images.jpg", "label": 0}
```

### Format for InfoNCE

```json lines
# LLM
{"query": "sentence1", "response": "sentence2"}
# MLLM
{"query": "<image>", "response": "sentence", "images": "/some/images.jpg"}
{"query": "<image>sentence1", "response": "<image>sentence2", "rejected_response": ["<image>sentence1", "<image>sentence2"], "images": ["/some/images.jpg", "/some/images.jpg", "/some/images.jpg", "/some/images.jpg"]}
```

InfoNCE loss supports the following environment variables:
1. `INFONCE_TEMPERATURE`: The temperature parameter. If not set, the default value is 0.01.
2. `INFONCE_USE_BATCH`: Determines whether to use `rejected_response` within the sample (hard negative samples) or to use all `responses` within a batch. The default is `True`, which means using responses within the batch.
3. `INFONCE_HARD_NEGATIVES`: The number of hard negatives. If not set, all samples in `rejected_response` will be used. Since the lengths may not be consistent, a for loop will be used to compute the loss (which is slower). If set to a specific number, and there are not enough samples, the missing number will be randomly sampled. If there are excess samples, the first `INFONCE_HARD_NEGATIVES` will be selected.
4. `INFONCE_MASK_FAKE_NEGATIVE`: Masks out fake negatives. The default is set to False. When enabled, it checks if a sample's similarity is greater than the positive sample's similarity plus 0.1. If so, the sample's similarity is set to -inf to prevent the leakage of the positive sample.

> It is also possible to set the number of hard negatives to be equal in the dataset, so that even if not set, the for loop method will not be used, thereby speeding up computation.
>
> `rejected_response` can also be omitted. In this case, `INFONCE_USE_BATCH` remains `True` and will use other samples within the batch as rejected responses.

The evaluation of InfoNCE loss includes the following metrics:
- mean_neg: The average of all hard negatives
- mean_pos: The average of all positives
- margin: The average of (positive - max hard negative)

## Scaffolding

SWIFT provides two scaffold training scripts:

- [GTE Model](https://github.com/tastelikefeet/swift/blob/main/examples/train/embedding/train_gte.sh)
- [GME Model](https://github.com/tastelikefeet/swift/blob/main/examples/train/embedding/train_gme.sh)

## Inference

SWIFT has supported the deployment of GME、GTE、Qwen3-Embedding models，please check[here](https://github.com/modelscope/ms-swift/blob/main/examples/deploy/embedding/client.py).

You can also use the original model's code for inference:

https://www.modelscope.cn/models/iic/gte_Qwen2-7B-instruct

https://www.modelscope.cn/models/iic/gme-Qwen2-VL-7B-Instruct

If you've used other models to train embedding from scratch (for example, the original `qwen2-vl` model + `--task_type embedding`), you can also use gme's inference code, but please note:

https://www.modelscope.cn/models/iic/gme-Qwen2-VL-7B-Instruct/file/view/master/gme_inference.py?status=1#L111

Please modify the template here to match the model's own template to ensure the final embeddings align correctly. It's particularly important to note that the template for the gme model is different from the chatml template for the `qwen2-vl` or `qwen2.5-vl` series. In its inference code, the ending character is `<|endoftext|>` rather than `<|im_end|>`.