ogma-base  ·  13.3M efficient text embedding model  ·  MTEB 57.02

High-quality English text embedding model for semantic search, RAG, vector search, retrieval, clustering, classification, STS, and agent memory — MTEB 57.02, 13.3M parameters, 1024-token context

Ogma Base is the quality-first mid-size model in the Ogma family. At 13.3M parameters it scores 57.02 MTEB in our canonical 66-task Ogma paper results, using only 59% of MiniLM-L6-v2's parameters and handling 4× longer input sequences (1024 vs 256 tokens). The sweet spot for quality-first RAG pipelines and agent memory.

Why the name Ogma?

Ogma is named after Ogma (also written Oghma), the Irish god associated with eloquence and credited in myth with inventing Ogham, an early alphabet for encoding language into symbols. That is the core job of an embedding model: turn language into compact vectors that machines can search, compare, cluster, and reason over.

Use cases

ogma-base is the quality-first Ogma model for semantic search, RAG retrieval, agent memory, vector databases, document retrieval, text classification, clustering, STS / sentence similarity, and retrieval-heavy agent pipelines. It is aimed at users who want better quality than MiniLM-class models while keeping the model small enough for practical CPU deployment.

Good fits:

• Production RAG and enterprise search where retrieval quality matters but the model still needs to be lightweight.
• Local or private embedding services for teams that want to avoid external embedding APIs for sensitive text.
• Agent memory systems with long context chunks, symmetric query/document encoding (SYM everywhere, or QRY/QRY), and frequent retrieval.
• Efficient CPU deployments where 13.3M parameters is easier to host than larger embedding transformers.
• Classification, clustering, and routing features where embedding quality directly affects downstream decisions.

Choose ogma-base when you want the strongest general-purpose Ogma model before moving into larger, accuracy-first territory.

Highlights

• 🏆 MTEB avg 57.02 — canonical Ogma paper result over 66/66 MTEB English tasks
• 📏 1024-token context — 4× longer than all-MiniLM-L6-v2 (256 tokens)
• 🔀 Symmetric routing via task tokens — encode everything with [SYM], or use [QRY]/[QRY] for retrieval (queries and documents both encoded with task="qry"); benchmark both routes on your task
• 📐 Matryoshka dims: [256, 128, 64, 32] — one model, any precision
• 🛡️ +4.0% F1 on prompt injection detection vs MiniLM (same architecture series)

Performance

MTEB English — 66/66 tasks (category-averaged)

Benchmarked with MTEB v2.10.7 on the standard 66-task English benchmark using category averaging (same methodology as the MTEB leaderboard).

Why choose Ogma Base?

ogma-base is the recommended choice when you want a strong quality/size tradeoff without going to full transformer scale (bge, e5). It is MiniLM-class on the external comparison line while being smaller and context-aware.

Safety — Toxicity & Prompt Injection Detection

Evaluated on the Ogma transformer architecture (same family). Embeddings are extracted then fed to a logistic regression (LR) or MLP classifier head — the embedding model itself is not fine-tuned. Evaluated against all-MiniLM-L6-v2 as baseline.

1. Jigsaw Toxic Comment Classification

Dataset: Arsive/toxicity_classification_jigsaw — Binary toxicity classification Train: 25,960 · Test: 6,490

Ogma (LR) leads MiniLM (LR) by +2.01% F1. MiniLM (MLP) leads on this dataset — the additional training data (25K samples) allows the MLP to compensate for MiniLM's slightly weaker base representations.

2. Prompt Injection Detection — deepset/prompt-injections

Dataset: deepset/prompt-injections — Binary injection detection Train: 546 · Test: 116 (low-data regime)

Ogma leads across both classifiers: +4.03% F1 (MLP), +4.23% F1 (LogReg). Ogma's representations are better separated in the low-data regime — it achieves 100% precision with LogReg, meaning zero false positives.

3. Prompt Injection Detection — neuralchemy/Prompt-injection-dataset

Dataset: neuralchemy/Prompt-injection-dataset — Binary injection detection Train: 4,391 · Test: 942

Ogma leads across all metrics: +0.78% F1 (MLP), +0.55% F1 (LR). Both models perform well at scale; Ogma maintains its edge and achieves higher AUC-ROC (99.37% vs 98.92%).

Summary

Ogma is a stronger feature extractor for prompt injection detection — the safety-critical task for agent pipelines. MiniLM edges ahead on toxicity when given sufficient labelled data and a more powerful classifier head. For agentic use cases where detecting adversarial instructions is the priority, Ogma representations are the better choice.

Architecture

Key design choices:

• Task token prepend: A learnable task token ([QRY], [DOC], or [SYM]) is prepended to the input sequence before the transformer. Recommended inference route: [QRY]/[QRY] — encode both queries and documents with [QRY]; this benchmarked highest on MTEB. [SYM] everywhere is the next-best symmetric alternative. We do not recommend [DOC] at inference time — it is exposed for downstream fine-tuning, not as an asymmetric query/document route.
• Matryoshka training: The model is trained with Matryoshka Representation Learning, meaning embeddings truncated to any supported sub-dimension remain well-calibrated without retraining.
• Mean pooling: The average of all token outputs (excluding padding) produces the sentence embedding, which consistently outperforms CLS-token pooling in the Ogma architecture family.
• L2 normalisation: All outputs are unit-normalised; cosine similarity == dot product == euclidean similarity (up to a constant), simplifying downstream usage.

Usage

Installation

pip install torch tokenizers transformers huggingface_hub

Basic Encoding

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("axiotic/ogma-base", trust_remote_code=True).eval()
tok   = AutoTokenizer.from_pretrained("axiotic/ogma-base", trust_remote_code=True)

sentences = [
    "The quick brown fox jumps over the lazy dog",
    "A fast auburn vulpine leaps over an idle canine",
    "The capital of France is Paris",
]
emb = model.embed(sentences, task="sym", tokenizer=tok)
# emb.shape → (256,) per sentence, L2-normalised

sim = (emb[0] @ emb[1]).item()   # cosine sim == dot product (L2-normalised)
print(f"paraphrase: {sim:.4f}")

task="sym" is a safe default for all similarity tasks (STS, clustering, classification) and for retrieval. Ogma is trained for symmetric routing — queries and documents are always encoded with the same task token. The two recommended routes are:

1. [SYM] for everything (the safe default above), or
2. [QRY]/[QRY] — encode both queries and documents with task="qry".

Try both on your downstream task; either can win depending on the data, and [QRY]/[QRY] is the natural starting point when fine-tuning a classifier or retrieval head on top of the embeddings.

Retrieval

Encode queries and documents with the same task token. Below we show the [QRY]/[QRY] route — both calls use task="qry". This is intentional (Ogma is symmetric, not asymmetric); swap in task="sym" to compare the SYM route on your data.

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("axiotic/ogma-base", trust_remote_code=True).eval()
tok   = AutoTokenizer.from_pretrained("axiotic/ogma-base", trust_remote_code=True)

queries = ["What is knowledge distillation?"]
docs = [
    "Knowledge distillation trains a smaller student model to mimic a larger teacher.",
    "The Eiffel Tower is in Paris, France.",
]

q = model.embed(queries, task="qry", tokenizer=tok)   # (256,) per query  — symmetric: both sides use qry
d = model.embed(docs,    task="qry", tokenizer=tok)   # (256,) per doc    — not a typo; Ogma is symmetric

scores = (q @ d.T).squeeze(0)   # cosine sim (L2-normalised, dot == cosine)
print(scores.tolist())          # [higher, lower] — first doc is relevant

Matryoshka — Flexible Dimensionality

Ogma is trained with Matryoshka Representation Learning. Slice and re-normalise to any supported sub-dimension with no retraining:

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

model = AutoModel.from_pretrained("axiotic/ogma-base", trust_remote_code=True).eval()
tok   = AutoTokenizer.from_pretrained("axiotic/ogma-base", trust_remote_code=True)

emb = model.embed(["hello world"], task="sym", tokenizer=tok)   # full 256d

for d in model.config.matryoshka_dims:
    sub = F.normalize(emb[:, :d], dim=-1)
    print(f"{d}d  norm={sub.norm(dim=-1).item():.4f}")

Model Family

All Ogma: MTEB 2.10.7, 66-task standard English set, category-averaged. MiniLM/Potion: published scores from the Model2Vec results page.

Training Details

Limitations

• No text generation. Ogma is an encoder-only embedding model.
• English only. Training data and evaluation are English-only.
• Slower than static models. Transformer inference is 40-100× slower than static models (Potion, Model2Vec) on CPU. The trade-off: contextual understanding and 4× longer sequences.
• Non-commercial licence. Due to distillation from a CC-BY-NC-4.0 teacher, Ogma inherits the NonCommercial restriction. Commercial use requires a separate Jina AI licence or retraining with a permissive teacher (Apache 2.0 compatible models like BGE or E5 can substitute at the cost of a full retraining run).
• Reranking gap. Ogma lags behind MiniLM-L6-v2 on reranking tasks (category avg delta: -6.8). This is an architectural characteristic: the model optimises for semantic similarity and classification over pairwise ranking.

Licence & Attribution

This model is released under CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0 International).

Required attribution (must be included in all uses):

This model was trained via knowledge distillation from jina-embeddings-v5-text-small (https://huggingface.co/jinaai/jina-embeddings-v5-text-small) by Jina AI, licensed under CC-BY-NC-4.0.

Citation

@misc{ogma2026,
  title     = {Ogma: Efficient Dense Retrieval via Structured Embeddings},
  author    = {Axiotic AI},
  year      = {2026},
  url       = {https://huggingface.co/axiotic/ogma-base},
}

MTEB(eng, v2) — full 41-task results

Added 2026-07 — measured on the current 41-task MTEB(eng, v2) benchmark (subprocess-per-task harness, validated by reproducing minishlab/potion-base-8M at 0.5328 vs its official 53.33). The small-mteb numbers above use a 20-task subset and are not directly comparable.