EMD123/Otzaria-Embedding-V1-Flash-0.6B

עברית — לחצו כאן

Otzaria Embedding V1 Flash 0.6B

This is an advanced embedding model designed for semantic retrieval and search within Jewish and Torah texts.

Base Model

Qwen/Qwen3-Embedding-0.6B

Intended Use

Semantic search and information retrieval for:

• Modern Hebrew
• Rabbinic Hebrew
• Yeshivish and Torah Hebrew
• Biblical Hebrew
• Aramaic and Talmudic-style texts

Important Limitation

This model retrieves relevant passages. It is not a halachic authority and must not be used as a standalone halachic decisor.

Pooling Method Used During Training

Use last-token pooling — taking the vector of the final non-padding token — followed by L2 normalization.

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

model_id = "EMD123/Otzaria-Embedding-V1-Flash-0.6B"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=True
)

model = AutoModel.from_pretrained(
    model_id,
    trust_remote_code=True
)

model.eval()


def last_token_pool(last_hidden_states, attention_mask):
    if attention_mask[:, -1].sum() == attention_mask.shape[0]:
        return last_hidden_states[:, -1]

    sequence_lengths = attention_mask.sum(dim=1) - 1

    return last_hidden_states[
        torch.arange(last_hidden_states.shape[0]),
        sequence_lengths
    ]


texts = [
    "מה מברכים על במבה?",
    "על חטיף העשוי מתירס... ברכתו שהכל"
]

inputs = tokenizer(
    texts,
    padding=True,
    truncation=True,
    return_tensors="pt"
)

with torch.no_grad():
    outputs = model(**inputs)

    embeddings = last_token_pool(
        outputs.last_hidden_state,
        inputs["attention_mask"]
    )

    embeddings = F.normalize(
        embeddings,
        p=2,
        dim=-1
    )

print(embeddings @ embeddings.T)

License

The base model, Qwen/Qwen3-Embedding-0.6B, is distributed under the Apache License 2.0. All rights granted by the Apache License 2.0 with respect to the base model remain unaffected.

To the extent permitted by applicable law and by the base model license, the fine-tuned modifications and this merged model distribution are made available under the Otzaria Non-Commercial Model License 1.0.

The model may be used, copied, modified, and redistributed solely for personal, academic, educational, research, charitable, or other non-commercial purposes.

The following uses require prior written permission from the model publisher:

• Use in a commercial product or service
• Use by or on behalf of a for-profit organization
• Paid access to the model or to a service powered by it
• Use in an advertising-supported or revenue-generating service
• Commercial redistribution, resale, or sublicensing
• Training, fine-tuning, distillation, or development of another model intended for commercial use

Redistributions must preserve:

• This README
• The repository's LICENSE file
• Attribution to the base model
• All applicable Apache License 2.0 notices
• A clear notice describing any modifications made

These restrictions do not limit anyone's rights to obtain and use the original, unmodified Qwen base model independently under the Apache License 2.0.

See the LICENSE file in this repository for the complete license terms.

---

עברית

מודל ההטמעה Otzaria Embedding V1 Flash 0.6B

זהו מודל הטמעה מתקדם, המיועד לאחזור ולחיפוש סמנטי בתוך טקסטים יהודיים ותורניים.

מודל הבסיס

Qwen/Qwen3-Embedding-0.6B

השימוש המיועד

חיפוש סמנטי ואחזור מידע עבור:

• עברית מודרנית
• עברית רבנית
• עברית ישיבתית ותורנית
• עברית מקראית
• ארמית וטקסטים בסגנון תלמודי

מגבלה חשובה

המודל מאחזר קטעים רלוונטיים. הוא אינו סמכות הלכתית, ואין להשתמש בו לבדו כפוסק הלכה.

שיטת ה־Pooling שבה נעשה שימוש באימון

יש להשתמש ב־Last-token pooling — לקיחת הווקטור של הטוקן האחרון שאינו טוקן ריפוד — ולאחר מכן לבצע נרמול L2.

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

model_id = "EMD123/Otzaria-Embedding-V1-Flash-0.6B"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=True
)

model = AutoModel.from_pretrained(
    model_id,
    trust_remote_code=True
)

model.eval()


def last_token_pool(last_hidden_states, attention_mask):
    if attention_mask[:, -1].sum() == attention_mask.shape[0]:
        return last_hidden_states[:, -1]

    sequence_lengths = attention_mask.sum(dim=1) - 1

    return last_hidden_states[
        torch.arange(last_hidden_states.shape[0]),
        sequence_lengths
    ]


texts = [
    "מה מברכים על במבה?",
    "על חטיף העשוי מתירס... ברכתו שהכל"
]

inputs = tokenizer(
    texts,
    padding=True,
    truncation=True,
    return_tensors="pt"
)

with torch.no_grad():
    outputs = model(**inputs)

    embeddings = last_token_pool(
        outputs.last_hidden_state,
        inputs["attention_mask"]
    )

    embeddings = F.normalize(
        embeddings,
        p=2,
        dim=-1
    )

print(embeddings @ embeddings.T)

רישיון

מודל הבסיס, Qwen/Qwen3-Embedding-0.6B, מופץ תחת רישיון Apache 2.0. כל הזכויות שהוענקו ביחס למודל הבסיס מכוח רישיון Apache 2.0 נשארות בתוקפן.

ככל שהדבר מותר על פי הדין החל ועל פי רישיון מודל הבסיס, השינויים שנוצרו באמצעות הכוונון וההפצה הממוזגת של מודל זה זמינים תחת רישיון המודל הלא־מסחרי של אוצריא, גרסה 1.0.

מותר להשתמש במודל, להעתיקו, לשנותו ולהפיצו מחדש למטרות אישיות, אקדמיות, חינוכיות, מחקריות, צדקה או למטרות לא־מסחריות אחרות בלבד.

השימושים הבאים מחייבים קבלת אישור מראש ובכתב ממפרסם המודל:

• שימוש במוצר או בשירות מסחרי
• שימוש בידי ארגון למטרות רווח או מטעמו
• מתן גישה בתשלום למודל או לשירות המופעל באמצעותו
• שימוש בשירות הנתמך בפרסומות או מפיק הכנסות
• הפצה מסחרית, מכירה או מתן רישיון משנה בתשלום
• אימון, כוונון, זיקוק או פיתוח של מודל אחר המיועד לשימוש מסחרי

כל הפצה מחדש חייבת לשמור על:

• קובץ README זה
• קובץ LICENSE של המאגר
• ייחוס למודל הבסיס
• כל הודעות הרישוי הרלוונטיות של Apache 2.0
• הודעה ברורה המתארת את השינויים שבוצעו

הגבלות אלה אינן מצמצמות את זכותו של אדם להשיג את מודל הבסיס המקורי והבלתי־משונה של Qwen ולהשתמש בו בנפרד בהתאם לרישיון Apache 2.0.

תנאי הרישיון המלאים מופיעים בקובץ LICENSE שבמאגר. :::

LICENSE

Hugging Face מנחה שבשימוש ב־license: other יש להוסיף את נוסח הרישיון המלא בקובץ LICENSE ולהגדיר לו license_name. ([Hugging Face][2])

Otzaria Non-Commercial Model License 1.0

Copyright (c) 2026 Otzaria Embedding Model Publisher

1. DEFINITIONS

“Base Model” means Qwen/Qwen3-Embedding-0.6B and any components derived directly from it that remain governed by the Apache License, Version 2.0.

“Model” means the fine-tuned weights, adapter-derived modifications, merged model distribution, configuration files, and related materials distributed under this repository, excluding any rights in the Base Model that are granted independently under the Apache License, Version 2.0.

“Non-Commercial Use” means use that is not primarily intended for or directed toward commercial advantage, monetary compensation, revenue generation, or the operations of a for-profit organization.

“Commercial Use” includes, without limitation:

a. use in or for a commercial product or service;

b. use by or on behalf of a for-profit organization;

c. paid access to the Model or to any product, API, application, or service powered by the Model;

d. use in an advertising-supported or otherwise revenue-generating product or service;

e. resale, commercial redistribution, paid sublicensing, or rental of the Model;

f. use of the Model to train, fine-tune, distill, improve, or develop another model intended for Commercial Use; and

g. internal business use connected with commercial operations, product development, customer service, or revenue-generating activities.

2. NON-COMMERCIAL LICENSE GRANT

Subject to compliance with this License, the Licensor grants you a non-exclusive, worldwide, non-transferable, royalty-free license to use, reproduce, study, evaluate, modify, and redistribute the Model solely for Non-Commercial Use.

3. PERMITTED USES

Permitted uses include personal use, academic research, education, non-commercial experimentation, charitable activity, and use by non-profit organizations, provided that the use is not carried out for commercial advantage or monetary compensation.

4. COMMERCIAL USE

Commercial Use is prohibited unless the Licensor has granted prior written permission under a separate commercial license.

No rights for Commercial Use are granted by this License.

5. REDISTRIBUTION AND MODIFICATION

You may redistribute the Model or modifications of the Model solely for Non-Commercial Use, provided that you:

a. include a complete copy of this License;

b. preserve this repository’s README and all applicable copyright, attribution, patent, trademark, and license notices;

c. preserve all notices required by the Apache License, Version 2.0;

d. clearly state that the Model has been modified and describe the material modifications;

e. do not represent the modified Model as an official release of the original publisher; and

f. apply terms that are at least as restrictive as this License to your modifications and redistribution.

6. BASE MODEL RIGHTS

The Base Model is licensed separately under the Apache License, Version 2.0.

Nothing in this License is intended to revoke, limit, replace, or otherwise interfere with rights granted directly by the Base Model’s licensor under the Apache License, Version 2.0.

For the avoidance of doubt, a person may obtain and use the original, unmodified Base Model independently under its original license.

This License applies only to the extent that the Licensor owns or is legally entitled to license the fine-tuned modifications and the Model distributed through this repository.

7. ATTRIBUTION

Any public use, publication, demonstration, or redistribution of the Model must include reasonable attribution to:

“Otzaria Embedding V1 Flash 0.6B”

and to the Base Model:

“Qwen/Qwen3-Embedding-0.6B”

8. NO TRADEMARK LICENSE

This License does not grant permission to use the names, trademarks, logos, or branding of Otzaria, Qwen, Alibaba, or any other party, except as reasonably necessary for attribution.

9. TERMINATION

Your rights under this License terminate automatically if you violate any of its terms.

Upon termination, you must cease using and distributing the Model and delete all copies under your control, except to the extent that retention is required by law.

Sections concerning attribution, disclaimers, limitation of liability, and Base Model rights survive termination.

10. DISCLAIMER OF WARRANTY

THE MODEL IS PROVIDED “AS IS,” WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING, WITHOUT LIMITATION, WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, ACCURACY, RELIABILITY, NON-INFRINGEMENT, OR FITNESS FOR HALACHIC, LEGAL, MEDICAL, FINANCIAL, OR OTHER PROFESSIONAL DECISION-MAKING.

11. LIMITATION OF LIABILITY

TO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE LAW, THE LICENSOR SHALL NOT BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, CONSEQUENTIAL, EXEMPLARY, OR OTHER DAMAGES ARISING FROM THE MODEL OR FROM ITS USE, MODIFICATION, OR DISTRIBUTION.

12. COMMERCIAL LICENSING

Commercial licenses or permissions may be requested directly from the Model publisher.

13. ENTIRE LICENSE

This License constitutes the complete terms granted by the Licensor for the Model, except for rights in third-party components governed by their respective licenses.