GigaChat 3B Bidirectional Embedding Model

Базовая LLM: собственная предобученная модель с архитектурой Qwen3 (36 слоёв, hidden 2048, 16 attention-голов / 8 KV-голов, head_dim 128), self-attention сделан двунаправленным (encoder-style)
Тип пулинга: Mean pooling (усреднение)
Размерность эмбеддинга: 2048
Параметры: ~3B (веса в формате bfloat16)

Следующая итерация серии Giga-Embeddings. Модель текстовых эмбеддингов на основе архитектуры Qwen3, адаптированная под двунаправленное (encoder-style) внимание и обученная с контрастивной функцией потерь (InfoNCE). Модель строит плотные эмбеддинги предложений/абзацев для задач поиска (retrieval), семантического сравнения, классификации и кластеризации, показывая высокое качество на русском и английском языках.

Пулинг и нормализация

Модель обучалась с mean pooling + L2-нормализацией. Использование CLS/last-token пулинга даст неверные результаты. Если вы используете transformers напрямую, необходимо самостоятельно усреднить (mean-pool) по не-паддинговым токенам и затем применить L2-нормализацию (см. пример ниже). В примерах для sentence-transformers и vLLM это делается автоматически. Сравнивайте эмбеддинги через косинусную близость (скалярное произведение нормированных векторов).

Инструктивность

Модель обучалась в инструктивном стиле: для retrieval и других асимметричных задач необходимо добавлять инструкцию к запросу (query), а документы кодируются как есть, без инструкции. Формат:

Instruct: {описание задачи}
Query: {ваш текст}

Для симметричных задач (STS, дедупликация) можно использовать общую инструкцию либо не использовать её вовсе. Инструкцию выбирают под конкретную задачу — единственного «правильного» промпта не существует. Важно отметить, что инструкцию нужно добавлять только перед запросом, а не перед документом.

FAQ

Нужно ли добавлять инструкции к запросу?

Для асимметричных задач (retrieval) — да, добавьте к запросу инструкцию из одного предложения, описывающую задачу. Документы кодируются как есть, без инструкции. Для симметричных задач (STS, дедупликация) можно использовать общую инструкцию либо не использовать её вовсе.

Какой пулинг использовать?

Mean pooling (усреднение) по не-паддинговым токенам с последующей L2-нормализацией. Использование CLS/last-token пулинга даст неверные результаты.

Почему мои воспроизведённые результаты немного отличаются от указанных в карточке модели?

Разные версии библиотек transformers и pytorch могут вызывать незначительные, но ненулевые различия в результатах.

GigaChat 3B Bidirectional Embedding Model

Base LLM: self-pretrained model with Qwen3 architecture (36 layers, hidden 2048, 16 attention heads / 8 KV heads, head_dim 128), self-attention made bidirectional (encoder-style)
Pooling Type: Mean pooling
Embedding Dimension: 2048
Parameters: ~3B (weights are bfloat16)

The next iteration of the Giga-Embeddings series. A text embedding model based on the Qwen3 architecture, adapted for bidirectional (encoder-style) attention and trained with a contrastive (InfoNCE) objective. It produces dense sentence/passage embeddings for retrieval, semantic similarity, classification and clustering, with strong Russian and English performance.

Pooling & normalization (important)

This model was trained with mean pooling + L2 normalization. Using CLS/last-token pooling will give wrong results. If you use transformers directly, you must mean-pool over non-padding tokens yourself and then L2-normalize (see the example below). The sentence-transformers and vLLM examples do this for you. Compare embeddings with cosine similarity (dot product of normalized vectors).

Instructions / prompts

The model was trained in the instruction style: for retrieval and other asymmetric tasks, prepend a task instruction to the query (documents are embedded raw). The format is:

Instruct: {task description}
Query: {your text}

For symmetric tasks (STS, deduplication) you can either use a generic instruction or none at all. Choose the instruction per task; there is no single "correct" prompt.

FAQ

Do I need to add instructions to the query?

For asymmetric tasks (retrieval), yes — prepend a one-sentence task instruction to the query. Documents are embedded raw, without an instruction. For symmetric tasks (STS, deduplication) you can use a generic instruction or none at all.

Which pooling should I use?

Mean pooling over non-padding tokens, followed by L2 normalization. Using CLS/last-token pooling will give wrong results.

Why are my reproduced results slightly different from those reported?

Different versions of the transformers and pytorch libraries can cause small but non-zero differences in results.

Metrics*

Benchmark	old 3b	Giga-Embeddings-instruct-3B-0826	Giga-Embeddings-instruct-10B-A1.8B-0826
MTEB (rus)	74.16	74.57	74.99
MTEB (eng)	71.07	71.93	72.23
MTEB (code)	62.37	76.93	78.40
MTEB (multilingual)	55.51	63.9	65.60

Model / backend	512 tok	1024 tok	2048 tok	throughput vs 10B-A1.8B
Giga-Embeddings-instruct-3B-0826 / vLLM	87.9k/s	91.5k/s	90.4k/s	0.8x
Giga-Embeddings-instruct-10B-A1.8B-0826 / vLLM	112.6k/s	114.5k/s	102.3k/s	1.0x
Nemotron 8B / vLLM	42.6k/s	43.2k/s	41.7k/s	0.38x
Qwen3 Embedding 4B / vLLM	70.1k/s	73.2k/s	71.2k/s	0.64x
F2LLM-v2-8B / vLLM	43.2k/s	43.4k/s	42.6k/s	0.38x
NV-Embed-v2 / Transformers	25.6k/s	26.2k/s	25.6k/s	0.23x

* All metrics were measured on an H100 GPU with a batch size of 16.

Usage

Sentence Transformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "ai-sage/Giga-Embeddings-instruct-3B-0826",
    trust_remote_code=True,   # needed for the bidirectional modeling code
)

instruction = "Given a query, retrieve relevant passages"
queries = [f"Instruct: {instruction}\nQuery: Где столица России?"]
documents = ["Москва — столица Российской Федерации.",
             "Париж — столица Франции."]

q_emb = model.encode(queries, normalize_embeddings=True)
d_emb = model.encode(documents, normalize_embeddings=True)
print(model.similarity(q_emb, d_emb))

Transformers (manual mean pooling)

import torch
import torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer

path = "ai-sage/Giga-Embeddings-instruct-3B-0826"
tok = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
model = AutoModel.from_pretrained(path, trust_remote_code=True,
                                  dtype=torch.bfloat16).cuda().eval()

def encode(texts):
    enc = tok(texts, return_tensors="pt", padding=True, truncation=True, max_length=512)
    enc = {k: v.cuda() for k, v in enc.items()}
    with torch.no_grad():
        hidden = model(**enc).last_hidden_state
    mask = enc["attention_mask"].unsqueeze(-1).to(hidden.dtype)
    emb = (hidden * mask).sum(1) / mask.sum(1).clamp(min=1e-6)   # mean pool
    return F.normalize(emb, dim=-1)                              # L2 normalize

instr = "Given a query, retrieve relevant passages"
q = encode([f"Instruct: {instr}\nQuery: Где столица России?"])
d = encode(["Москва — столица Российской Федерации.", "Париж — столица Франции."])
print((q @ d.T).cpu())

vLLM

vLLM serves this model as an embedding model using its native Qwen3 implementation. Bidirectional attention is enabled with is_causal=false; no custom code is required on the vLLM side.

from vllm import LLM
from vllm.config import PoolerConfig

llm = LLM(
    model="ai-sage/Giga-Embeddings-instruct-3B-0826",
    runner="pooling",
    convert="embed",
    hf_overrides={"is_causal": False, "architectures": ["Qwen3ForCausalLM"]},
    pooler_config=PoolerConfig(pooling_type="MEAN", use_activation=True),
    trust_remote_code=True,
)

instr = "Given a query, retrieve relevant passages"
outs = llm.encode([f"Instruct: {instr}\nQuery: Где столица России?",
                   "Москва — столица Российской Федерации."],
                  pooling_task="embed")
embs = [o.outputs.data for o in outs]

Or via the OpenAI-compatible server:

vllm serve ai-sage/Giga-Embeddings-instruct-3B-0826 \
    --runner pooling --convert embed \
    --hf-overrides '{"is_causal": false, "architectures": ["Qwen3ForCausalLM"]}' \
    --override-pooler-config '{"pooling_type": "MEAN", "use_activation": true}' \
    --trust-remote-code

* this example is for latest vllm=0.26.0 release, for older vllm versions you might need to change pooler config argument from use_activation to normalize

SGLang

Support comes from PR #35531 – [Model] Add qwen3 bidirectional embedding, which adds the Qwen3BidirectionalModel architecture used by ai-sage/Giga-Embeddings-instruct-3B-0826.

Source branch: feat/qwen3-bidirectional-embedding on fork Lossfull/sglang
Pinned commit: 604a3634d235b11dcf4abd4bc012cfa1f7bde43b
The change is pure Python (no CUDA/kernel rebuild needed).

Once the PR is merged this whole guide collapses to "use a recent SGLang release." Until then, use one of the two methods below.

1. Official SGLang Docker + apply the PR patch (recommended)

The official image already ships SGLang as an editable install with all CUDA kernels prebuilt. The PR is pure Python, so you just patch the files in place — nothing is compiled, and the Python source stays matched to the image's kernels. This is the method that was verified end-to-end for this guide.

# 1. Start the verified image (its SGLang is editable at /sgl-workspace/sglang).
docker run --gpus all -it --shm-size 16g \
  -p 30000:30000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --entrypoint /bin/bash \
  lmsysorg/sglang:nightly-dev-20260818-c0b6474b

# --- everything below runs INSIDE the container ---

# 2. Download the PR diff.
curl -fL -H "Accept: application/vnd.github.v3.diff" \
  -o /tmp/pr35531.diff \
  https://api.github.com/repos/sgl-project/sglang/pulls/35531

# 3. Apply it onto the image's editable source tree (takes effect immediately).
cd /sgl-workspace/sglang
git apply -v /tmp/pr35531.diff        # or: patch -p1 < /tmp/pr35531.diff

# 4. Sanity check: the new architecture must resolve to the native class.
python3 -c "from sglang.srt.models.registry import ModelRegistry; \
c,a=ModelRegistry.resolve_model_cls('Qwen3BidirectionalModel'); \
print('OK:', a, '->', c.__module__)"
# Expect: OK: Qwen3BidirectionalModel -> sglang.srt.models.qwen3_embedding

2. Build from source (no Docker)

Use this on a bare CUDA machine (or a plain PyTorch container). It builds the matching kernels, so it is heavier but fully self-contained.

# Clone the PR branch (or the exact commit).
git clone https://github.com/Lossfull/sglang.git
cd sglang
git checkout feat/qwen3-bidirectional-embedding
# Optional: pin the exact reviewed commit
# git checkout 604a3634d235b11dcf4abd4bc012cfa1f7bde43b

# Install SGLang + all runtime deps (compiles/pulls sgl-kernel, flashinfer, ...).
pip install --upgrade pip
pip install -e "python[all]"

3. Serve the model

Same command regardless of install method:

python3 -m sglang.launch_server \
  --model-path ai-sage/Giga-Embeddings-instruct-3B-0826 \
  --is-embedding \
  --trust-remote-code \
  --host 0.0.0.0 --port 30000

--is-embedding — serve as an embedding model (the arch is auto-classified as non-generative anyway, but this is explicit and safe).
--trust-remote-code — required (custom config class in the checkpoint).
Disabled CUDA graph / radix cache / chunked prefill are applied automatically — you don't set them.
Multi-GPU: add --tp-size N if you want to shard across GPUs.

The server is ready when you see: The server is fired up and ready to roll!

4. Test it

cURL

curl -s http://localhost:30000/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
        "model": "ai-sage/Giga-Embeddings-instruct-3B-0826",
        "input": "What is the capital of France?"
      }' | python3 -c "import sys,json; d=json.load(sys.stdin); \
e=d['data'][0]['embedding']; print('dim:', len(e), 'first5:', e[:5])"

Fine-tune guide

Finetuning Giga-Embeddings with ms-swift

This guide shows how to contrastively finetune the Giga-Embeddings models with ms-swift using an InfoNCE loss, for both LoRA and full-parameter training.

Covered models (HuggingFace):

Model	Size	Architecture
`ai-sage/Giga-Embeddings-instruct-480M-0826`	0.48B	Qwen3 bidirectional
`ai-sage/Giga-Embeddings-instruct-3B-0826`	3B	Qwen3 bidirectional
`ai-sage/Giga-Embeddings-instruct-10B-A1.8B-0826`	10B MoE (A1.8B)	DeepSeek-V3 bidirectional

All are mean-pooling SentenceTransformer models, so we load them through ms-swift's SentenceTransformersLoader.

Requirements

# ms-swift from main — the current PyPI release does not yet contain the
# SentenceTransformer full-parameter save fix.
pip install "git+https://github.com/modelscope/ms-swift.git"

# sentence-transformers pinned to 5.3.0.
pip install "sentence-transformers==5.3.0"

torch and transformers are pulled in automatically. A CUDA GPU is required for training.

For full-parameter finetuning of the 10B MoE model (multi-GPU, DeepSpeed ZeRO-3), also install DeepSpeed:

pip install deepspeed

1. Register the model

The Giga-Embeddings architectures aren't in ms-swift's built-in registry, so register them once and point them at SentenceTransformersLoader. Save this as custom_register.py:

# custom_register.py
from swift.model import Model, ModelGroup, ModelMeta, register_model
from swift.model.register import SentenceTransformersLoader
from swift.template import TemplateType

# Qwen3-bidirectional models (480M, 3B)
register_model(ModelMeta(
    'giga_embeddings',
    [ModelGroup([
        Model('ai-sage/Giga-Embeddings-instruct-480M-0826', 'ai-sage/Giga-Embeddings-instruct-480M-0826'),
        Model('ai-sage/Giga-Embeddings-instruct-3B-0826', 'ai-sage/Giga-Embeddings-instruct-3B-0826'),
    ])],
    SentenceTransformersLoader,
    template=TemplateType.dummy,
    architectures=['Qwen3BidirectionalModel'],
))

# DeepSeek-V3-bidirectional MoE model (10B-A1.8B)
#
# Under DeepSpeed ZeRO-3, MoE routing makes different ranks run different expert
# submodules, which breaks ZeRO-3's per-parameter cross-rank coordination
# ("Detected a disagreement on list length between rank0 and rankN"). The fix is to
# mark the MoE block as a ZeRO-3 leaf module. We do it in a small loader subclass
# (harmless when ZeRO-3 / DeepSpeed isn't used — only needed for the 10B on ZeRO-3).
class GigaMoESentenceTransformersLoader(SentenceTransformersLoader):
    def get_model(self, model_dir, config, processor, model_kwargs):
        model = super().get_model(model_dir, config, processor, model_kwargs)
        try:
            from deepspeed.utils import set_z3_leaf_modules
            set_z3_leaf_modules(model, ['DeepseekV3MoE'])
        except Exception:
            pass
        return model

register_model(ModelMeta(
    'giga_embeddings_moe',
    [ModelGroup([
        Model('ai-sage/Giga-Embeddings-instruct-10B-A1.8B-0826', 'ai-sage/Giga-Embeddings-instruct-10B-A1.8B-0826'),
    ])],
    GigaMoESentenceTransformersLoader,
    template=TemplateType.dummy,
    architectures=['DeepseekV3BidirectionalModel'],
))

Pass it to any swift sft command with --custom_register_path custom_register.py.

2. Prepare your dataset

One JSON object per line. Each row is a query with one positive document and any number of hard negatives:

{"messages": [{"role": "user", "content": "What is the capital of France?"}],
 "positive_messages": [[{"role": "assistant", "content": "Paris is the capital of France."}]],
 "negative_messages": [[{"role": "assistant", "content": "Berlin is the capital of Germany."}],
                       [{"role": "assistant", "content": "The Eiffel Tower is a landmark."}]]}

messages — the query / anchor.
positive_messages — list of positive documents (≥ 1).
negative_messages — list of hard negatives (optional but recommended).

ms-swift lays each row out as anchor + positive + negatives with labels [1, 0, 0, …], which is what the InfoNCE loss consumes.

If your retrieval task uses an instruction prefix, prepend it to the query text (the models were trained with Instruct: <task>\nQuery: <query>).

3. Train

The examples use these InfoNCE environment variables (they are read from the environment, not passed as CLI flags):

export INFONCE_TEMPERATURE=0.05
export INFONCE_USE_BATCH=True       # in-batch negatives (see note below)
export INFONCE_HARD_NEGATIVES=7     # hard negatives per query (match your data)

In-batch negatives (INFONCE_USE_BATCH): keep True for general retrieval data where each query has a distinct positive. Set it to False if your dataset has a small set of shared positive documents (e.g. many queries mapping to the same handful of answers) — otherwise another query's positive becomes a false negative for yours and hurts training. When False, rely on the curated hard negatives.

LoRA

swift sft \
  --custom_register_path custom_register.py \
  --model_type giga_embeddings \
  --model ai-sage/Giga-Embeddings-instruct-480M-0826 \
  --use_hf true \
  --task_type embedding \
  --loss_type infonce \
  --tuner_type lora \
  --lora_rank 8 --lora_alpha 32 \
  --dataset ./train.jsonl \
  --split_dataset_ratio 0.0 \
  --max_length 512 \
  --num_train_epochs 1 \
  --per_device_train_batch_size 8 \
  --learning_rate 1e-4 \
  --torch_dtype bfloat16 \
  --attn_impl sdpa \
  --logging_steps 5 --save_steps 500 \
  --output_dir ./output

Full parameter

Same as above with --tuner_type full and a lower learning rate. For the 480M and 3B this fits comfortably on a single 80 GB GPU:

swift sft \
  --custom_register_path custom_register.py \
  --model_type giga_embeddings \
  --model ai-sage/Giga-Embeddings-instruct-3B-0826 \
  --use_hf true \
  --task_type embedding --loss_type infonce \
  --tuner_type full \
  --dataset ./train.jsonl --split_dataset_ratio 0.0 \
  --max_length 512 --num_train_epochs 1 \
  --per_device_train_batch_size 4 --learning_rate 1e-5 \
  --torch_dtype bfloat16 --attn_impl sdpa \
  --logging_steps 5 --save_steps 500 \
  --output_dir ./output

Multi-GPU

For multiple GPUs, launch with NPROC_PER_NODE. Plain DDP (no ZeRO) works well and fits full-parameter finetuning of the 3B on 8×80 GB:

NPROC_PER_NODE=8 CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 swift sft \
  --custom_register_path custom_register.py \
  --model_type giga_embeddings \
  --model ai-sage/Giga-Embeddings-instruct-3B-0826 \
  --use_hf true \
  --task_type embedding --loss_type infonce \
  --tuner_type full \
  --dataset ./train.jsonl --split_dataset_ratio 0.0 \
  --max_length 2048 --num_train_epochs 1 \
  --per_device_train_batch_size 4 \
  --learning_rate 1e-5 --warmup_ratio 0.03 --lr_scheduler_type cosine \
  --gradient_checkpointing true \
  --torch_dtype bfloat16 --attn_impl sdpa \
  --dataloader_num_workers 4 --dataset_num_proc 8 \
  --logging_steps 5 --save_steps 500 --save_total_limit 3 \
  --output_dir ./output

Tips:

--dataset_num_proc N parallelizes tokenization (helps for large datasets).
With in-batch negatives on, they are gathered across all GPUs, giving a larger effective negative pool as you add GPUs.

4. The 10B MoE model

Use --model_type giga_embeddings_moe and --model ai-sage/Giga-Embeddings-instruct-10B-A1.8B-0826. Everything above applies, plus:

Memory. Full-parameter finetuning of the 10B does not fit on a single 80 GB GPU with standard AdamW. Use multi-GPU DeepSpeed ZeRO-3:

NPROC_PER_NODE=8 swift sft \
  --custom_register_path custom_register.py \
  --model_type giga_embeddings_moe \
  --model ai-sage/Giga-Embeddings-instruct-10B-A1.8B-0826 \
  --use_hf true \
  --task_type embedding --loss_type infonce \
  --tuner_type full \
  --deepspeed zero3 \
  --gradient_checkpointing true \
  --dataset ./train.jsonl --split_dataset_ratio 0.0 \
  --max_length 2048 --num_train_epochs 1 \
  --per_device_train_batch_size 2 --learning_rate 1e-5 \
  --torch_dtype bfloat16 --attn_impl sdpa \
  --logging_steps 5 --save_steps 500 \
  --output_dir ./output

(Requires pip install deepspeed.) LoRA on the 10B fits on a single 80 GB GPU without DeepSpeed — use --tuner_type lora as in §3.

Expert routing. ms-swift's embedding trainer does not add a MoE load-balancing auxiliary loss. For long full-parameter runs, monitor expert utilization.

5. Verify a finetuned checkpoint

Reload with sentence-transformers and confirm positives score higher than negatives. A correct full-parameter checkpoint reloads with no "missing keys" warnings:

from sentence_transformers import SentenceTransformer

m = SentenceTransformer("./output/<run>/checkpoint-<N>", trust_remote_code=True, device="cuda")
emb = m.encode(
    ["What is the capital of France?",
     "Paris is the capital of France.",
     "Berlin is in Germany."],
    convert_to_tensor=True, normalize_embeddings=True,
)
print("cos(query, positive) =", float(emb[0] @ emb[1]))  # should be clearly higher
print("cos(query, negative) =", float(emb[0] @ emb[2]))

Full-parameter output is a complete SentenceTransformer checkpoint — load the output directory directly.
LoRA output is an adapter. Load the base model and apply the adapter, or merge first with swift export --adapters ./output/<run>/checkpoint-<N> --merge_lora true.

InfoNCE options reference

Set via environment variables:

Variable	Default	Meaning
`INFONCE_TEMPERATURE`	`0.1`	Softmax temperature (lower = sharper).
`INFONCE_USE_BATCH`	`True`	Use in-batch (and cross-GPU) negatives.
`INFONCE_HARD_NEGATIVES`	–	Hard negatives kept per query.
`INFONCE_MASK_FAKE_NEGATIVE`	`False`	Mask in-batch negatives scoring above the positive (guards against false negatives).
`INFONCE_INCLUDE_QQ` / `INFONCE_INCLUDE_DD`	`False`	Add query-query / doc-doc terms to the denominator (Qwen3-Embedding style).

Key swift sft flags:

Flag	Meaning
`--task_type embedding`	Enable embedding training (ST pooling + embedding trainer).
`--loss_type infonce`	InfoNCE contrastive loss.
`--tuner_type lora` / `full`	LoRA (default) vs full-parameter.
`--use_hf true`	Resolve `--model` from the HuggingFace Hub.
`--custom_register_path`	Path to `custom_register.py`.
`--split_dataset_ratio 0.0`	No automatic validation split.

Troubleshooting

AttributeError: 'NoneType' object has no attribute 'items' during save — you're on the PyPI release of ms-swift; install from main (see Requirements).
Reloaded model gives base-model quality / "missing keys" on load after full finetuning — your sentence-transformers is newer than 5.3; pin ==5.3.0.
Model downloads from ModelScope instead of HuggingFace (or is not found) — add --use_hf true.

Paper