ArieLLL123/Meivin-Embed-0.1.1

Meivin Embed 0.1.1

Meivin – מודל הטמעה סמנטי תורני

Meivin Embed 0.1.1 is an experimental semantic embedding model designed for semantic search and reranking of Judaic texts.

This release is an incremental improvement over Meivin Embed 0.1.

It is intentionally released as 0.1.1, rather than 0.2, because it represents an experimental improvement and evaluation step rather than a major new model generation.

What changed from 0.1?

Version 0.1.1 was trained to improve the model's ability to rank relevant Judaic passages higher within a candidate result pool.

Training

• Training queries: 1,659
• Validation queries: 207
• Test queries: 201
• Negative examples: 4 per training example
• Effective batch size: 16
• Epochs: 3
• Learning rate: 2e-5
• Query maximum length: 128
• Passage maximum length: 512

The best checkpoint was selected using validation results.

Training loss

Epoch	Average loss
1	1.1622
2	0.3316
3	0.1695

The selected checkpoint is Epoch 1.

Evaluation

Test reranking

Metric	Baseline	Meivin Embed 0.1.1
NDCG@10	0.4765	0.9401
MRR@10	0.3809	0.9863
Success@1	0.2587	0.9751
Precision@10	0.2149	0.2637

These results show a substantial improvement in the model's ability to rerank the evaluated candidate pools.

Important limitation

These numbers do not represent full-corpus retrieval performance.

The evaluation reranks a fixed candidate pool for each query. It does not measure recall from the entire Seforim corpus, nor does it represent the performance of a newly rebuilt hybrid search index.

Therefore, the high NDCG, MRR and Success@1 values should be interpreted specifically as reranking results.

The repository contains the evaluation files used during development.

Repository contents

• best/ — selected model checkpoint
• last/ — final training checkpoint
• evaluation.json — final evaluation results
• baseline_validation.json — baseline validation results
• test_rankings.json — test ranking data
• history.json — training history

The best/ directory contains the model backbone, tokenizer and projection layer required for inference.

Status

Experimental — Meivin Embed 0.1.1

This release is intended for experimentation, evaluation and further development.

It should not yet be considered a stable production release or the final Meivin architecture.

License

This model is distributed under the Personal Use License, Version 1.0.

The license permits private, personal and non-commercial use only. Public use, distribution, organizational use and incorporation into products, services or other systems require prior written permission from the copyright holder.

See LICENSE for the complete license terms.

Attribution to the Otzaria Project (אוצריא) is required for permitted use.

Attribution

Otzaria Project (אוצריא)

https://otzaria.org