/train-sentence-transformers

Model Architectures (SentenceTransformer)

The SentenceTransformer class is a torch.nn.Sequential of modules. The common shape is Transformer + Pooling (+ optional Normalize / Dense), but four distinct architecture families are supported and the right choice depends on the task.

The four architecture families

Below, each family with concrete setup.

Encoder models (the default)

The historical default and still usually the right choice for text embeddings.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("microsoft/mpnet-base")
# Auto-constructs: Transformer(feature-extraction) -> Pooling(mean).

When SentenceTransformer("<checkpoint>") is called with a raw HF encoder, it auto-wraps the transformer and adds Pooling(..., pooling_mode="mean").

To customize pooling or add modules:

from sentence_transformers import SentenceTransformer
from sentence_transformers.sentence_transformer.modules import Normalize, Pooling, Transformer

transformer = Transformer("answerdotai/ModernBERT-base")
pooling = Pooling(transformer.get_embedding_dimension(), pooling_mode="cls")    # or "mean", "lasttoken", ...
model = SentenceTransformer(modules=[transformer, pooling, Normalize()])

Pooling modes:

• mean (default): average of token embeddings, masked to the attention mask. Strongest default.
• cls: embedding of the [CLS] token. Works if the base was CLS-pretrained.
• max: element-wise max across tokens. Rare.
• mean_sqrt_len_tokens: mean scaled by √seq_len. Empirically helps on some tasks.
• weightedmean: token-position-weighted mean. Useful for decoder bases as a non-last-token alternative.
• lasttoken: embedding of the last token. Required for causal-LM bases (see decoder section below).

Don't switch pooling mid-training. Pick once.

Decoder / causal LLM models

Strong at long context, instruction following, multilingual. Memory-hungry, typically LoRA-trained rather than full fine-tuned.

Two setup paths depending on whether the model was already adapted for embeddings:

# Path A: already-adapted embedding checkpoint (ships with the right modules):
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("Qwen/Qwen3-Embedding-0.6B")   # just works

# Path B: raw decoder LLM, build the pipeline manually:
from sentence_transformers import SentenceTransformer
from sentence_transformers.sentence_transformer.modules import Normalize, Pooling, Transformer

transformer = Transformer(
    "Qwen/Qwen2.5-0.5B",
    transformer_task="text-generation",         # critical: causal attention, no bidirectional
    processor_kwargs={"padding_side": "left"},  # last-token pooling wants left-padding
)
pooling = Pooling(transformer.get_embedding_dimension(), pooling_mode="lasttoken")
model = SentenceTransformer(modules=[transformer, pooling, Normalize()])

Skipping transformer_task="text-generation" or pooling_mode="lasttoken" on a raw decoder gives embeddings that look plausible until you benchmark.

Why last-token pooling: causal attention means only the last token has seen the full sequence. Mean-pooling a causal model averages embeddings that only saw prefixes. The result doesn't represent the whole input.

For training decoder bases:

• Learning rate: typically 1e-4 or higher (not 2e-5 like encoders).
• LoRA is almost always the right choice for >1B-param bases. See ../scripts/train_sentence_transformer_with_lora_example.py (its docstring covers when to use, hyperparams, QLoRA for 7B+, and adapter sharing).

Static embeddings

StaticEmbedding skips the transformer entirely. Each token maps to a pre-computed vector via a lookup table. No attention, no contextualization.

When to use:

• CPU inference, no GPU, browser / edge / on-device deployment.
• Need <10MB model size.
• Latency budget <1ms per embedding.
• Have >1M training pairs (contextualization is replaced by per-token optimization, which takes data).

When NOT to use:

• Task needs contextual understanding (polysemy, syntax, long-range dependencies).
• You have <100k training pairs. The model won't learn enough.

Setup:

from sentence_transformers import SentenceTransformer
from sentence_transformers.sentence_transformer.modules import StaticEmbedding
from tokenizers import Tokenizer

tokenizer = Tokenizer.from_pretrained("google-bert/bert-base-uncased")
static_embedding = StaticEmbedding(tokenizer, embedding_dim=512)
model = SentenceTransformer(modules=[static_embedding])

Train with MultipleNegativesRankingLoss on a large contrastive dataset (1M+ pairs).

Warm starts vs. random init: with >1M training samples, random-init beats StaticEmbedding.from_model2vec(...) or .from_distillation(...) warm starts. With smaller datasets, warm starts help.

# For smaller datasets (<100k), warm-start:
static_embedding = StaticEmbedding.from_model2vec("minishlab/potion-base-8M")
# or:
static_embedding = StaticEmbedding.from_distillation("sentence-transformers/all-MiniLM-L6-v2", vocabulary=list(tokenizer.get_vocab().keys()))

See ../scripts/train_sentence_transformer_static_embedding_example.py for a runnable end-to-end recipe (random init + MNRL + Matryoshka + bf16 + lr=2e-1) and the Static Embeddings blog post for benchmarks.

Multimodal via VLM backbone

Modern vision-language models can be loaded directly and produce joint text+image embeddings:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "Qwen/Qwen3-VL-Embedding-2B",
    model_kwargs={"attn_implementation": "flash_attention_2"},  # do NOT set torch_dtype here; see training_args.md
    processor_kwargs={"min_pixels": 28 * 28, "max_pixels": 600 * 600},
)

# Check which modalities this model supports:
print(model.modalities)
# ['text', 'image', 'video', 'message']

Training data can mix text, PIL images, image paths/URLs, audio, and mixed-modality dicts like {"image": <PIL>, "text": "describe this"}. The data collator handles preprocessing via the model's preprocess method.

Install multimodal extras: pip install "sentence-transformers[image]" (or [audio], [video]).

Precision: load in fp32 and pass bf16=True (or fp16=True) to TrainingArguments. Autocast handles the inference path. Don't set torch_dtype="bfloat16" in model_kwargs: it puts Adam state in bf16 and silently degrades quality (see training_args.md).

Multimodal via Router

Instead of one VLM backbone, compose separate encoders per modality:

from sentence_transformers import SentenceTransformer
from sentence_transformers.sentence_transformer.modules import Dense, Pooling, Router, Transformer

# Text encoder
text_encoder = Transformer("sentence-transformers/all-MiniLM-L6-v2")
text_pooling = Pooling(text_encoder.get_embedding_dimension(), pooling_mode="mean")
# Project text to match image encoder's dim
text_projection = Dense(text_encoder.get_embedding_dimension(), 768)

# Image encoder (SigLIP outputs pooled embeddings directly)
image_encoder = Transformer("google/siglip2-base-patch16-224")

router = Router(
    sub_modules={
        "text": [text_encoder, text_pooling, text_projection],
        "image": [image_encoder],
    },
)
model = SentenceTransformer(modules=[router])

Warning: Router-based models have unaligned embedding spaces at init. You must train to align them. Use a Dense projection layer when dimensions differ. Task-based routing (different encoders for queries vs. documents) is also supported via route_mappings. See the Router docstring.

Gotchas

• Decoder base with mean pooling: silently produces garbage embeddings. Always use lasttoken.
• Router multimodal without training: the separate encoders' embedding spaces are unaligned at init. Don't expect useful cross-modal similarity until you've trained with a loss that aligns the spaces.
• StaticEmbedding with fewer than 100k pairs: the model won't learn enough. Either warm-start via from_model2vec / from_distillation, or use a regular encoder.
• Large VLM backbones on consumer GPUs: combine LoRA + attn_implementation="flash_attention_2". With LoRA only, you can additionally pass torch_dtype="bfloat16". The bf16 base weights are frozen, so the Adam-state precision concern from the precision rule above doesn't apply (the LoRA adapter stays fp32, so its optimizer state stays fp32). Without LoRA, follow the precision rule: keep weights fp32 and rely on bf16=True autocast.