aria-embedding-311m-en-de

English + German retrieval embedding model, finetuned from ibm-granite/granite-embedding-311m-multilingual-r2 (ModernBERT, 311M parameters, 32,768-token context, 768-dim embeddings with Matryoshka truncation to 512/256/128).

The finetune targets general EN+DE retrieval with an emphasis on long documents: on an 11-task EN/DE suite it improves the base model's overall nDCG@10 from 0.587 to 0.605, ahead of snowflake-arctic-embed-m-v2.0 (0.601) at comparable size, with the largest gains on long-document and German legal retrieval.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("4thel00z/aria-embedding-311m-en-de")

queries = [
    "Who is Miss Delmer?",
    "Wann ist eine Erlaubnis zu widerrufen?",
]
documents = ["...", "..."]

query_embeddings = model.encode(queries, normalize_embeddings=True)
document_embeddings = model.encode(documents, normalize_embeddings=True)
scores = query_embeddings @ document_embeddings.T

No query or passage prefixes are required (same convention as the base model). For smaller embeddings, load with truncate_dim=512, 256, or 128.

Evaluation

All numbers are nDCG@10, computed in a single reproducible harness with per-model encoding configurations, so every model below is measured under identical conditions. They are in-harness results, not official leaderboard numbers.

Task groups: de = GermanQuAD, GerDaLIR; en-short = NanoMSMARCO, NanoNQ, NanoHotpotQA; longembed = the six LongEmbed tasks, where needle and passkey are means over eight context lengths (256 to 32,768 tokens) on a shared mixed-length corpus.

model	params	de	en-short	longembed	overall
aria-embedding-311m-en-de	311M	0.575	0.647	0.594	0.605
snowflake-arctic-embed-m-v2.0	305M	0.565	0.766	0.470	0.601
bge-m3	568M	0.556	0.711	0.518	0.595
granite-embedding-311m-multilingual-r2 (base)	311M	0.543	0.647	0.571	0.587
multilingual-e5-large	560M	0.506	0.732	0.372	0.537

Per task, against the base model:

task	base	aria	delta
GermanQuAD	0.930	0.936	+0.006
GerDaLIR	0.157	0.213	+0.057
NanoMSMARCO	0.545	0.550	+0.005
NanoNQ	0.658	0.656	-0.002
NanoHotpotQA	0.737	0.735	-0.002
LongEmbed 2WikiMQA	0.849	0.852	+0.003
LongEmbed NarrativeQA	0.511	0.580	+0.069
LongEmbed QMSum	0.471	0.480	+0.009
LongEmbed SummScreenFD	0.957	0.969	+0.012
LongEmbed Needle (mean of 8 lengths)	0.224	0.215	-0.009
LongEmbed Passkey (mean of 8 lengths)	0.417	0.467	+0.050

Training

Two stages of full finetuning with sentence-transformers, both wrapping every loss in MatryoshkaLoss (768/512/256/128) at 8,192-token training sequence length; the model keeps the base's 32,768-token inference context.

1. Multi-task epoch. KL-divergence distillation of BAAI/bge-reranker-v2-m3 logits (teacher temperature 3.0) over four mined hard negatives per query on MS MARCO (EN) and mMARCO (DE), 100k pairs each, alongside contrastive training (CachedMultipleNegativesRankingLoss, scale 50) on GermanDPR, GerDaLIR with mined hard negatives, and synthetic long-document EN/DE pairs.
2. NarrativeQA top-up. One contrastive epoch on 32,593 (question, full story) pairs from the NarrativeQA train split. Train stories are disjoint from the LongEmbed test stories, and every training query was additionally checked against the evaluation query sets by normalized text and dropped on overlap.

Limitations

• Short-text English retrieval trails snowflake-arctic-embed-m-v2.0 (0.647 vs 0.766 on the NanoBEIR group); pick this model when German and/or long documents matter.
• Needle-in-a-haystack style exact-phrase retrieval degrades beyond ~2k tokens of context, as in the base model.
• Trained and evaluated on English and German; other languages inherit whatever the base model provides, untested.

Attribution

Base model by IBM Granite (Apache-2.0). Distillation teacher: BAAI/bge-reranker-v2-m3. Training data: MS MARCO, mMARCO, GermanDPR, GerDaLIR, NarrativeQA, and synthetic pairs generated from FinePDFs documents.