Artemis Embed v1
Artemis Embed v1 is an English general-purpose dense embedding model built on
answerdotai/ModernBERT-base.
It produces one L2-normalized dense vector per text and is intended for semantic similarity, retrieval, document search, clustering, and downstream classification features.
Architecture
- Backbone:
answerdotai/ModernBERT-base - Pooling: masked mean
- Full embedding dimension: 768
- Matryoshka dimensions: 768 / 512 / 256 / 128
- Normalization: L2 normalization after truncation
- Tested maximum sequence length in the current release: 128 tokens
- Current training recipe: LoRA + contrastive training + hard negatives + Matryoshka learning
Usage
from sentence_transformers import SentenceTransformer
import torch.nn.functional as F
import torch
model = SentenceTransformer("Omarbm52/Artemis-Embed-v1")
texts = [
"A dog is running in a park.",
"A puppy runs outside.",
]
emb = model.encode(
texts,
convert_to_tensor=True,
normalize_embeddings=True,
)
# Full 768D embeddings
emb_768 = emb
# Example Matryoshka 256D embeddings.
emb_256 = F.normalize(emb[:, :256], p=2, dim=-1)
When using a truncated Matryoshka dimension, truncate the prefix first and then L2-normalize again.
Internal development results
These numbers come from the project's internal development suite and are not a final held-out MTEB result.
| Metric | 768D |
|---|---|
| Retrieval Recall@1 | 0.8675 |
| Retrieval Recall@5 | 0.9700 |
| Retrieval MRR | 0.909534 |
| STS Spearman | 0.345035 |
| Classification accuracy | 0.872 |
| Clustering NMI | 0.589730 |
| Internal aggregate | 0.702423 |
Matryoshka development results
| Dimension | R@1 | R@5 | MRR | STS Spearman |
|---|---|---|---|---|
| 768 | 0.8675 | 0.9700 | 0.909534 | 0.345035 |
| 512 | 0.8550 | 0.9650 | 0.901461 | 0.344076 |
| 256 | 0.8275 | 0.9525 | 0.883264 | 0.330863 |
| 128 | 0.8250 | 0.9450 | 0.876263 | 0.329936 |
Limitations
- English-only v1.
- Current release metrics are internal development metrics.
- Final MTEB English v2 evaluation remains open.
- The raw ModernBERT baseline is not included in the exported development result table.
- Training arms were not all compute-controlled, so the current winner should not be interpreted as evidence that LoRA is inherently better than full fine-tuning.
- Dataset licensing, overlap, and contamination review must be completed before a research-grade final release claim.
Matryoshka note
Truncating from 768 dimensions to a smaller prefix reduces downstream vector storage, bandwidth, similarity-computation cost, and index size. It does not reduce the ModernBERT backbone forward-pass FLOPs.