You can deploy this model and start earning money today!

Views

No views yet

ogma-micro · 2.3M efficient text embedding model · MTEB 52.18

Name: ogma-micro
Availability: InStock

Ultra-small English text embedding model for semantic search, RAG, vector search, clustering, classification, and agent memory — MTEB 52.18, 2.3M parameters, 128d output

Maintained successor (MIT): the permissively licensed model in this family — including the CoreML build that used to sit under coreml/ here — is axiotic/open-ogma-micro. This repository's own weights remain CC-BY-NC-4.0 and it no longer hosts that artifact.

Ogma Micro is the most compact model in the Ogma family. At 2.3M parameters and 8.9 MB it scores 52.18 MTEB in our 66-task run while staying small enough to ship in browsers and on-device runtimes. Outputs 128-dimensional embeddings for maximum indexing efficiency. For extreme latency, edge, and browser workloads.

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-micro is the smallest Ogma model, built for on-device embedding, edge search, browser-side retrieval, local semantic search, agent memory, deduplication, classification, clustering, and privacy-sensitive applications where sending text to an external embedding API is undesirable.

Good fits:

Mobile and desktop apps that need local text embeddings without a large model download.
Browser, WebAssembly, and extension-style workflows where package size and vector index size matter.
Serverless and high-fanout applications that need many cheap embedding calls with predictable memory use.
Local-first search over notes, messages, logs, support tickets, snippets, or small document collections.
Efficient vector databases where 128-dimensional embeddings reduce storage, bandwidth, and ANN latency.

Choose ogma-micro when footprint matters more than absolute benchmark quality. Move up to ogma-mini or ogma-small when you can spend more memory for stronger representations.

Highlights

🏆 MTEB avg 52.18 — compact 2.3M-parameter model from the canonical Ogma paper results
📦 8.9 MB — smallest in the family
📐 128-dim output — half the index size of other Ogma models
📏 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: [128, 64, 32] — compress to 32d for ultra-low memory indexing

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).

Category	ogma-micro	all-MiniLM-L6-v2	Δ vs MiniLM
Classification	59.53	62.62	-3.09
Clustering	36.88	41.94	-5.06
PairClassification	78.62	82.37	-3.75
Reranking	49.74	58.04	-8.30
Retrieval	33.09	41.95	-8.86
STS	75.63	78.90	-3.27
Summarization	31.77	30.81	+0.96
Overall	52.18	56.09	-3.91

Why choose Ogma Micro?

ogma-micro is for when you need the absolute smallest possible model that still achieves competitive MTEB scores. Note the 128-dim output — your vector index will be half the size of other Ogma models. Use ogma-mini if you can afford 3.5M parameters.

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

Model	Classifier	Accuracy	F1	Precision	Recall	AUC-ROC
Ogma	LogReg	89.12%	88.26%	89.09%	87.44%	95.74%
Ogma	MLP	88.91%	87.98%	89.14%	86.85%	95.92%
MiniLM	LogReg	87.32%	86.25%	87.46%	85.07%	94.96%
MiniLM	MLP	91.71%	91.24%	90.13%	92.39%	97.16%

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)

Model	Classifier	Accuracy	F1	Precision	Recall	AUC-ROC
Ogma	LogReg	86.21%	84.62%	100.0%	73.33%	97.77%
Ogma	MLP	90.52%	90.27%	96.23%	85.0%	98.1%
MiniLM	LogReg	82.76%	80.39%	97.62%	68.33%	94.52%
MiniLM	MLP	87.07%	86.24%	95.92%	78.33%	93.96%

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

Model	Classifier	Accuracy	F1	Precision	Recall	AUC-ROC
Ogma	LogReg	95.22%	95.93%	95.84%	96.01%	99.30%
Ogma	MLP	95.44%	96.16%	94.89%	97.46%	99.37%
MiniLM	LogReg	94.59%	95.38%	95.46%	95.29%	98.92%
MiniLM	MLP	93.95%	94.85%	94.59%	95.11%	98.92%

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

Task	Ogma best F1	MiniLM best F1	Δ
Jigsaw Toxicity	88.26% (LR)	91.24% (MLP)	−2.98%
deepset Injection	90.27% (MLP)	86.24% (MLP)	+4.03%
neuralchemy Injection	96.16% (MLP)	95.38% (LR)	+0.78%

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

Property	Value
Architecture	Custom Transformer
Internal dim (`d_model`)	128
Output dim (`d_output`)	128
Transformer layers	2
Attention heads	2
Vocabulary	30,000 (SentencePiece / AlbertTokenizer)
Max sequence length	1,024 tokens
Pooling	Mean pooling
Task tokens	`[QRY]` (query), `[DOC]` (document), `[SYM]` (symmetric)
Matryoshka dims	[32, 64, 128]
Output normalisation	L2 (unit sphere)
Parameters	2.3M
Model file	`model.safetensors` (8.9 MB)

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

python

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("axiotic/ogma-micro", trust_remote_code=True).eval()
tok   = AutoTokenizer.from_pretrained("axiotic/ogma-micro", 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 → (128,) 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:

[SYM] for everything (the safe default above), or
[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.

python

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("axiotic/ogma-micro", trust_remote_code=True).eval()
tok   = AutoTokenizer.from_pretrained("axiotic/ogma-micro", 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)   # (128,) per query  — symmetric: both sides use qry
d = model.embed(docs,    task="qry", tokenizer=tok)   # (128,) 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:

python

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

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

emb = model.embed(["hello world"], task="sym", tokenizer=tok)   # full 128d

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

Model	Params	Size	MTEB Avg	Class	Clust	PairClass	Rerank	Ret	STS	Summ	d_out	Context
ogma-large	32.4M	124 MB	57.41	68.6	41.6	84.0	53.1	43.7	83.7	30.9	256	1024
ogma-base	13.3M	51 MB	57.02	67.74	41.49	83.73	51.25	42.36	82.84	29.73	256	1024
ogma-small	8.6M	33 MB	56.32	66.49	40.69	82.91	50.51	42.05	82.00	29.59	256	1024
ogma-mini	3.5M	14 MB	53.06	61.77	37.38	79.66	47.39	36.21	77.71	31.33	256	1024
ogma-micro	2.3M	8.9 MB	52.18	59.53	36.88	78.62	49.74	33.09	75.63	31.77	128	1024
all-MiniLM-L6-v2	22.7M	87 MB	56.09	62.62	41.94	82.37	58.04	41.95	78.90	30.81	384	256
potion-base-32M	32.0M	123 MB	51.22	66.0	39.2	78.2	50.9	32.2	73.9	29.8	256	inf
potion-base-8M	7.6M	29 MB	50.03	64.44	32.93	76.62	49.73	31.71	73.24	29.28	256	inf

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

Training Details

Property	Value
Teacher model	`jinaai/jina-embeddings-v5-text-small` (CC-BY-NC-4.0)
Training paradigm	Knowledge distillation from cached teacher embeddings
Training data	~7M curated English sentence pairs
Tokenizer	AlbertTokenizer (SentencePiece, vocab=30,000)
Embedding initialisation	PCA of teacher embeddings (128d) projected to d_model
Loss	Distillation + contrastive (balanced schedule)
Evaluation framework	MTEB 2.10.7

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: -8.3). 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

bibtex

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

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.

mean	Class	Clust	PairClass	Rerank	Retrieval	STS	Summ
0.5427	0.6537	0.4058	0.7862	0.4244	0.3498	0.7525	0.2967