VespaEmbed

No-code training for embedding models. Train custom embedding models with a web UI or CLI.

Features

• Web UI - Visual interface for configuring and monitoring training
• CLI - Command-line interface for scripting and automation
• Multiple Tasks - Support for pairs, triplets, similarity scoring, and unsupervised learning
• Loss Variants - Choose from multiple loss functions per task
• Matryoshka Embeddings - Train multi-dimensional embeddings for flexible retrieval
• LoRA Support - Parameter-efficient fine-tuning with LoRA adapters
• Unsloth Integration - Faster training with Unsloth optimizations
• HuggingFace Integration - Load datasets, models from HuggingFace Hub, push models to Hub

Installation

pip install vespaembed

Optional Dependencies

# For Unsloth acceleration (requires NVIDIA/AMD GPU)
pip install vespaembed[unsloth]

Development Installation

git clone https://github.com/vespaai-playground/vespaembed.git
cd vespaembed
uv sync --extra dev

Quick Start

Web UI

Launch the web interface:

vespaembed

Open http://localhost:8000 in your browser. The UI lets you:

• Upload training data (CSV or JSONL)
• Select task type and base model
• Configure hyperparameters
• Monitor training progress
• Download trained models

CLI

Train a model from the command line:

vespaembed train \
  --data examples/data/pairs.csv \
  --task pairs \
  --base-model sentence-transformers/all-MiniLM-L6-v2 \
  --epochs 3

Or use a YAML config file:

vespaembed train --config config.yaml

Tasks

VespaEmbed supports 4 training tasks based on your data format:

Pairs

Text pairs for semantic search. Use when you have query-document pairs without explicit negatives.

Data format:

anchor,positive
What is machine learning?,Machine learning is a subset of AI...
How does photosynthesis work?,Photosynthesis converts sunlight...

Loss variants: mnr (default), mnr_symmetric, gist, cached_mnr, cached_gist

Triplets

Text triplets with hard negatives. Use when you have explicit negative examples.

Data format:

anchor,positive,negative
What is Python?,Python is a programming language...,A python is a large snake...

Loss variants: mnr (default), mnr_symmetric, gist, cached_mnr, cached_gist

Similarity

Text pairs with similarity scores (STS-style). Use when you have continuous similarity labels.

Data format:

sentence1,sentence2,score
A man is playing guitar,A person plays music,0.85
The cat is sleeping,A dog is running,0.12

Loss variants: cosine (default), cosent, angle

TSDAE

Unsupervised learning with denoising auto-encoder. Use when you only have unlabeled text for domain adaptation.

Data format:

text
Machine learning is transforming how we analyze data.
Natural language processing enables computers to understand human language.

Configuration

CLI Arguments

vespaembed train \
  --data <path>              # Training data (CSV, JSONL, or HF dataset)
  --task <task>              # Task type: pairs, triplets, similarity, tsdae
  --base-model <model>       # Base model name or path
  --project <name>           # Project name (optional)
  --eval-data <path>         # Evaluation data (optional)
  --epochs <n>               # Number of epochs (default: 3)
  --batch-size <n>           # Batch size (default: 32)
  --learning-rate <lr>       # Learning rate (default: 2e-5)
  --optimizer <opt>          # Optimizer (default: adamw_torch)
  --scheduler <sched>        # LR scheduler (default: linear)
  --matryoshka               # Enable Matryoshka embeddings
  --matryoshka-dims <dims>   # Dimensions (default: 768,512,256,128,64)
  --unsloth                  # Use Unsloth for faster training
  --subset <name>            # HuggingFace dataset subset
  --split <name>             # HuggingFace dataset split

Optimizers

Option	Description
adamw_torch	AdamW (default)
adamw_torch_fused	Fused AdamW (faster on CUDA)
adamw_8bit	8-bit AdamW (memory efficient)
adafactor	Adafactor (memory efficient, no momentum)
sgd	SGD with momentum

Schedulers

Option	Description
linear	Linear decay (default)
cosine	Cosine annealing
cosine_with_restarts	Cosine with warm restarts
constant	Constant learning rate
constant_with_warmup	Constant after warmup
polynomial	Polynomial decay

YAML Configuration

task: pairs
base_model: sentence-transformers/all-MiniLM-L6-v2

data:
  train: train.csv
  eval: eval.csv            # optional

training:
  epochs: 3
  batch_size: 32
  learning_rate: 2e-5
  warmup_ratio: 0.1
  weight_decay: 0.01
  fp16: true
  eval_steps: 500
  save_steps: 500
  logging_steps: 100
  optimizer: adamw_torch    # adamw_torch, adamw_8bit, adafactor, sgd
  scheduler: linear         # linear, cosine, constant, polynomial

output:
  dir: ./output
  push_to_hub: false
  hf_username: null

# Optional: LoRA configuration
lora:
  enabled: false
  r: 64
  alpha: 128
  dropout: 0.1
  target_modules: [query, key, value, dense]

# Optional: Matryoshka dimensions
matryoshka_dims: [768, 512, 256, 128, 64]

# Optional: Loss variant (uses task default if not specified)
loss_variant: mnr

HuggingFace Datasets

Load datasets directly from HuggingFace Hub:

vespaembed train \
  --data sentence-transformers/all-nli \
  --subset triplet \
  --split train \
  --task triplets \
  --base-model sentence-transformers/all-MiniLM-L6-v2

CLI Commands

Command	Description
vespaembed	Launch web UI (default)
vespaembed serve	Launch web UI
vespaembed train	Train a model
vespaembed evaluate	Evaluate a model
vespaembed export	Export model to ONNX
vespaembed info	Show task information

Output

Trained models are saved to ~/.vespaembed/projects/<project-name>/:

~/.vespaembed/projects/my-project/
├── final/              # Final trained model
├── onnx/               # ONNX export of the final model (model.onnx + tokenizer)
├── checkpoint-500/     # Training checkpoints
├── checkpoint-1000/
└── logs/               # TensorBoard logs

Column Aliases

VespaEmbed automatically recognizes common column name variations:

Task	Expected	Also Accepts
pairs	anchor	query, question, sent1, sentence1, text1
pairs	positive	document, answer, pos, sent2, sentence2, text2
triplets	negative	neg, hard_negative, sent3, sentence3, text3
similarity	sentence1	sent1, text1, anchor, query
similarity	sentence2	sent2, text2, positive, document
similarity	score	similarity, label, sim_score
tsdae	text	sentence, sentences, content, input

Important: Columns are matched by name (or alias), not by position. For example, with a pairs task:

• [anchor, positive] or [query, document] → works ✓
• [document, query] → still works (names identify roles, not position) ✓
• [foo, bar] → fails (no matching column names or aliases) ✗

Columns named score, scores, label, or labels (and aliases like similarity) are treated as labels/targets.

Development

# Run tests
uv run pytest tests/

# Run tests with coverage
uv run pytest tests/ --cov=vespaembed

# Format code
make format

# Lint
make lint

License

Apache 2.0