You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

Meivin Embed 0.1

Total downloads Parameters Formats

ื”ื˜ืžืขื•ืช ืกืžื ื˜ื™ื•ืช ืœื—ื™ืคื•ืฉ ื‘ืกืคืจื•ืช ืชื•ืจื ื™ืช ื‘ืขื‘ืจื™ืช ื•ื‘ืืจืžื™ืช โ€” ื‘ืžืฉืงืœื™ื ืจื’ื™ืœื™ื ื•ื‘ึพONNX ืœื”ืจืฆื” ืžืงื•ืžื™ืช.

Meivin Embed ืžืžื™ืจ ืฉืืœื•ืช ื•ืžืงื˜ืขื™ ื˜ืงืกื˜ ืœื•ื•ืงื˜ื•ืจื™ื ืฉืœ 256 ืžืžื“ื™ื, ื›ื“ื™ ืœืืคืฉืจ ื—ื™ืคื•ืฉ ืœืคื™ ืžืฉืžืขื•ืช ื’ื ื›ืืฉืจ ื ื™ืกื•ื— ื”ืฉืืœื” ืฉื•ื ื” ืžืœืฉื•ืŸ ื”ืžืงื•ืจ. ื”ื•ื ืžื™ื•ืขื“ ืœืžืคืชื—ื™ ืžื ื•ืขื™ ื—ื™ืคื•ืฉ ืชื•ืจื ื™ื™ื, ืกืคืจื™ื•ืช ื“ื™ื’ื™ื˜ืœื™ื•ืช ื•ืžืขืจื›ื•ืช ืื—ื–ื•ืจ ืžืงื•ืจื•ืช ืขื‘ื•ืจ RAG.

ื”ื—ื‘ื™ืœื” ื›ื•ืœืœืช ืžืฉืงืœื™ื ืจื’ื™ืœื™ื ื‘ึพSafetensors, ื’ืจืกืช ONNX FP32 ื•ื’ืจืกืช ONNX INT8 ืงื•ืžืคืงื˜ื™ืช. ืœืื—ืจ ื”ื•ืจื“ืช ื”ืงื‘ืฆื™ื ืืคืฉืจ ืœืงื•ื“ื“ ื˜ืงืกื˜ ืžืงื•ืžื™ืช ื‘ืืžืฆืขื•ืช ONNX Runtime. ื”ืžื•ื“ืœ ืžืคื™ืง ื•ืงื˜ื•ืจื™ื; ื™ืฆื™ืจืช ืื™ื ื“ืงืก, ื—ื™ืคื•ืฉ ื‘ื• ื•ื”ืคืงืช ืชืฉื•ื‘ื” ื˜ืงืกื˜ื•ืืœื™ืช ื ืขืฉื™ื ื‘ืจื›ื™ื‘ื™ื ืื—ืจื™ื ืฉืœ ื”ื™ื™ืฉื•ื.

ื›ืœ ืฉื™ืžื•ืฉ ืžืกื—ืจื™ ืืกื•ืจ ืœืœื ืื™ืฉื•ืจ ืžืคื•ืจืฉ ืžืจืืฉ ื•ื‘ื›ืชื‘ ืžืืจื™ืืœ ื“ื ื™ืืœื™ โ€” arieldaniely@gmail.com. ืืกื•ืจ ืœืฉืœื‘ ืื• ืœื”ืฉืชืžืฉ ื‘ืžื•ื“ืœ ื‘ื›ืœ ืฆื•ืจื” ื‘ืžืขืจื›ืช, ืชื•ื›ื ื” ืื• ืฉื™ืจื•ืช ืœืžื˜ืจื•ืช ืจื•ื•ื—, ื’ื ืื ื”ืจื›ื™ื‘ ืฉื‘ื• ื”ื•ื ืžื•ื˜ืžืข ื—ื™ื ืžื™. ืจืื• ื”ืจื™ืฉื™ื•ืŸ ื”ืžืœื.

The Meivin family | ืกื“ืจืช ืžื™ื™ื‘ื™ืŸ

Model Purpose Output
Meivin Embed 0.1 Semantic retrieval for Judaic libraries and RAG source retrieval L2-normalized 256-dimensional embeddings
Meivin MLM 0.1 Contextual masked-token prediction and language research Scores over a 32,000-token vocabulary

Built for the language of Jewish texts, with compact models that run locally. The two products share a documented training foundation and serve complementary tasks. This release is available under Meivin Non-Commercial License 1.0; commercial licensing and collaboration: Ariel Daniely.

English overview

A compact Hebrew and Aramaic encoder for semantic search in Judaic literature. It maps queries and passages to normalized 256-dimensional vectors, runs locally, and ships as complete PyTorch/Safetensors weights, ONNX FP32 and ONNX INT8. The regular checkpoint contains 42,260,993 parameters; the ONNX files are approximately 168.2 MB and 42.5 MB respectively. Use [QUERY] and [PASSAGE] prefixes with the included tokenizer. Full training report documents the training lineage and retrieval evaluations.

ื‘ืžื‘ื˜ ืื—ื“

ืžืืคื™ื™ืŸ ืขืจืš
ืชื—ื•ื ืกืคืจื•ืช ืชื•ืจื ื™ืช ื‘ืขื‘ืจื™ืช ื•ื‘ืืจืžื™ืช
ืคืœื˜ ื•ืงื˜ื•ืจ ืฉืœ 256 ืžืžื“ื™ื, ืžื ื•ืจืžืœ L2
ืคื•ืจืžื˜ Safetensors FP32; ONNX opset 17
ื’ืจืกืื•ืช PyTorch Safetensors FP32, โ€ONNX FP32, โ€ONNX INT8
ืคืจืžื˜ืจื™ื 42,260,993 ื‘ื—ื‘ื™ืœืช ื”ืžืฉืงืœื™ื ื”ืžืœืื”
ื’ื•ื“ืœ ืงื•ื‘ืฅ ื”ืžื•ื“ืœ FP32: ื›ึพ168.2 MB; INT8: ื›ึพ42.5 MB, ื‘ื™ื—ื™ื“ื•ืช ืขืฉืจื•ื ื™ื•ืช
ืงืœื˜ ืจืฆืฃ ืื—ื“ ื‘ื›ืœ ืงืจื™ืื”, ื‘ืื•ืจืš ืžืฉืชื ื”
ืชืงืจืช ื”ืืจื›ื™ื˜ืงื˜ื•ืจื” 512 ื˜ื•ืงื ื™ื, ื›ื•ืœืœ ืงื™ื“ื•ืžืช ื•ื˜ื•ืงื ื™ื ืžื™ื•ื—ื“ื™ื
ืื•ืจืš ื‘ื“ื•ื’ืžืช ื”ืฉื™ืžื•ืฉ 256 ื˜ื•ืงื ื™ื
ื’ื™ืฉื” ื‘ืงืฉืช ื’ื™ืฉื” ื•ืื™ืฉื•ืจ ื™ื“ื ื™ ื‘ึพHugging Face
ืจื™ืฉื™ื•ืŸ ืืจื˜ื™ืคืงื˜ื™ ื”ืžื•ื“ืœ Meivin Non-Commercial License 1.0; ืฉื™ืžื•ืฉ ืžืกื—ืจื™ ืจืง ื‘ืื™ืฉื•ืจ ื‘ื›ืชื‘

0.1 ื”ื™ื ื’ืจืกืช ื”ืžื•ืฆืจ ื”ืžืชื•ืขื“ืช ื‘ื›ืจื˜ื™ืก ื–ื”. ื”ืžืฉืงื•ืœื•ืช ืžื‘ื•ืกืกื•ืช ืขืœ checkpoint ื”ืื—ื–ื•ืจ Round 2; ื”ืžื™ืชื•ื’ ืื™ื ื• ืžืฆื™ื™ืŸ ืื™ืžื•ืŸ ื—ื“ืฉ. ื”ื“ื•ื’ืžืื•ืช ืžืฉืชืžืฉื•ืช ื‘ืฉื ื”ืžืื’ืจ ื”ืจืฉืžื™ Meivin-Embed-0.1.

ื”ืชื—ืœื” ืžื”ื™ืจื”

ื‘ืงืฉื• ื’ื™ืฉื” ื‘ึพืขืžื•ื“ ื”ืžื•ื“ืœ. ืœืื—ืจ ืื™ืฉื•ืจื”, ื”ืชื—ื‘ืจื• ืœื—ืฉื‘ื•ืŸ ื‘ืขืœ ื”ื”ืจืฉืื”:

pip install numpy onnxruntime tokenizers huggingface_hub
hf auth login

ื”ื“ื•ื’ืžื” ืžืงื•ื“ื“ืช ืฉืืœื” ื•ืฉื ื™ ืžืงื˜ืขื™ ื”ื“ื’ืžื” ื•ืžื“ืจื’ืช ืื•ืชื ืœืคื™ ื“ืžื™ื•ืŸ ืงื•ืกื™ื ื•ืก:

import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download
from tokenizers import Tokenizer

REPO = "ArieLLL123/Meivin-Embed-0.1"
MODEL_FILE = "seforim-embed-round2-int8.onnx"
import json
config_path = hf_hub_download(REPO, "config.json", token=True)
config = json.load(open(config_path, encoding="utf-8"))
model_path = hf_hub_download(REPO, MODEL_FILE, token=True)
tokenizer_path = hf_hub_download(REPO, "tokenizer.json", token=True)

tokenizer = Tokenizer.from_file(tokenizer_path)
tokenizer.enable_truncation(max_length=config["query_max_length"])
session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])

def encode(text: str, role: str) -> np.ndarray:
    prefixes = {"query": "[QUERY]", "passage": "[PASSAGE]"}
    tokens = tokenizer.encode(f"{prefixes[role]} {text}", add_special_tokens=True)
    inputs = {
        "input_ids": np.asarray([tokens.ids], dtype=np.int64),
        "attention_mask": np.asarray([tokens.attention_mask], dtype=np.int64),
    }
    return session.run(["embedding"], inputs)[0][0]

query = encode("ืžื” ืžืฉืžืขื•ืช ื”ื‘ื™ื˜ื•ื™ ื“ื‘ืจ ืฉืื™ื ื• ืžืชื›ื•ื•ืŸ?", "query")
passages = [
    "ื“ื‘ืจ ืฉืื™ื ื• ืžืชื›ื•ื•ืŸ ื”ื•ื ืžืขืฉื” ืฉืชื•ืฆืืชื• ื”ืืกื•ืจื” ืื™ื ื” ื›ื•ื•ื ืช ื”ืขื•ืฉื”.",
    "ืžืฆื•ื•ืช ืกืคื™ืจืช ื”ืขื•ืžืจ ื ื•ื”ื’ืช ื‘ื›ืœ ื™ื•ื ืžื™ืžื™ ื”ืกืคื™ืจื”.",
]
vectors = np.stack([encode(text, "passage") for text in passages])
scores = vectors @ query  # ื‘ื•ื•ืงื˜ื•ืจื™ื ืžื ื•ืจืžืœื™ื, ืžื›ืคืœื” ืกืงืœืจื™ืช ืฉื•ื•ื” ืœืงื•ืกื™ื ื•ืก
for index in np.argsort(-scores):
    print(f"{scores[index]:.4f}\t{passages[index]}")

ืœืฉื™ืžื•ืฉ ื‘ึพFP32, ืฉื ื• ืืช MODEL_FILE ืœึพseforim-embed-round2-fp32.onnx. ืงื•ื“ื“ื• ืฉืืœื•ืช ื•ืžืงื˜ืขื™ ืื™ื ื“ืงืก ื‘ืื•ืชื” ื’ืจืกืช ืงื•ื‘ืฅ ืžื•ื“ืœ ื•ื‘ืื•ืชื• ื˜ื•ืงื ื™ื™ื–ืจ.

Regular weights: PyTorch / Safetensors

model.safetensors contains the complete trained sentence encoder: BERT backbone, projection, LayerNorm and the stored training logit scale. A plain AutoModel loader is not sufficient for this custom sentence encoder: it would omit the retrieval head. Use the small explicit loader below after reviewing meivin_embed.py; no trust_remote_code=True is needed.

pip install torch transformers safetensors huggingface_hub numpy

Download the weights and loader at one pinned revision (see RELEASE_NOTES.md), then run:

from huggingface_hub import snapshot_download
import importlib.util
from transformers import AutoTokenizer

folder = snapshot_download("ArieLLL123/Meivin-Embed-0.1", token=True,
    allow_patterns=["config.json", "model.safetensors", "tokenizer*", "special_tokens_map.json", "meivin_embed.py"])
spec = importlib.util.spec_from_file_location("meivin_embed", folder + "/meivin_embed.py")
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)  # reviewed repository helper
model = module.MeivinEmbed.from_directory(folder)
tokenizer = AutoTokenizer.from_pretrained(folder)
query = module.encode(model, tokenizer, ["ืžื”ื• ื“ื‘ืจ ืฉืื™ื ื• ืžืชื›ื•ื•ืŸ?"], role="query")
passages = module.encode(model, tokenizer, ["ื“ื™ื ื™ ืฉื‘ืช ื•ืžืœืื›ื”", "ืกืคื™ืจืช ื”ืขื•ืžืจ"], role="passage")
print(passages @ query[0])

The explicit loader returns normalized vectors and supports padded batches. Regular weights retain gradients for further training; encode is an inference helper. Existing ONNX files retain their single-sequence input contract. PyTorch and ONNX FP32 were compared using short inputs and a 512-token input; see validation.

ื›ืœืœื™ ืฉื™ืœื•ื‘

  • ื”ื•ืกื™ืคื• [QUERY] ืœืคื ื™ ืฉืืœื•ืช ื•ึพ[PASSAGE] ืœืคื ื™ ืžืงื˜ืขื™ื, ืœืคื ื™ ื”ื˜ื•ืงื ื™ื–ืฆื™ื”. ืืœื” ื˜ื•ืงื ื™ื ืžื™ื•ื—ื“ื™ื ืฉื ืœืžื“ื• ื‘ืื™ืžื•ืŸ.
  • ื”ืฉืชืžืฉื• ื‘ื˜ื•ืงื ื™ื™ื–ืจ ื”ืžืฆื•ืจืฃ ื•ืฉืžืจื• ืขืœ ื”ื›ื ืช ื˜ืงืกื˜ ืขืงื‘ื™ืช ื‘ืื™ื ื“ื•ืงืก ื•ื‘ื—ื™ืคื•ืฉ.
  • ื”ืงืœื˜ื™ื input_ids ื•ึพattention_mask ื”ื ืžืกื•ื’ int64 ื•ื‘ืฆื•ืจื” [1, sequence_length]. ื”ื’ืจืฃ ืชื•ืžืš ื‘ืจืฆืฃ ืื—ื“ ื‘ื›ืœ ืงืจื™ืื”.
  • ื”ืคืœื˜ embedding ื”ื•ื float32[1, 256], ื’ื ื‘ึพINT8. ื”ึพpooling, ื”ื”ื˜ืœื” ื•ื”ื ืจืžื•ืœ ื›ืœื•ืœื™ื ื‘ื’ืจืฃ.
  • ื‘ื“ื•ื’ืžื” ืžื•ืคืขืœืช ืžื’ื‘ืœื” ืฉืœ 256 ื˜ื•ืงื ื™ื, ื‘ื“ื•ืžื” ืœื‘ืจื™ืจืช ื”ืžื—ื“ืœ ืฉืœ ื”ืžืงื•ื“ื“ ื”ืžืงื•ืžื™. ื”ืืจื›ื™ื˜ืงื˜ื•ืจื” ืžืืคืฉืจืช ืขื“ 512; ื‘ื“ืงื• ืื™ื›ื•ืช ื•ื”ืชืืžื” ืœืคื ื™ ืฉื™ื ื•ื™ ื”ืžื’ื‘ืœื”.
  • ื—ืœืงื• ืžืกืžื›ื™ื ืืจื•ื›ื™ื ืœืžืงื˜ืขื™ื ื•ืฉืžืจื• ื”ืคื ื™ื•ืช ืœืžืงื•ืจ ืœืฆื“ ื”ื•ื•ืงื˜ื•ืจื™ื. ื˜ืงืกื˜ ืžืขื‘ืจ ืœืžื’ื‘ืœื” ื ื—ืชืš.
  • ืฉื™ื ื•ื™ ื”ืžื•ื“ืœ, ื”ื›ื™ืžื•ืช, ื”ื˜ื•ืงื ื™ื™ื–ืจ ืื• ื”ื›ื ืช ื”ื˜ืงืกื˜ ืžื—ื™ื™ื‘ ื‘ื“ื™ืงืช ื”ืชืืžื” ื•ื‘ื“ืจืš ื›ืœืœ ืงื™ื“ื•ื“ ืžื—ื“ืฉ ืฉืœ ื”ืื™ื ื“ืงืก.

ืฆื™ื•ืŸ ื”ื“ืžื™ื•ืŸ ืžืฉืžืฉ ืœื“ื™ืจื•ื’ ืชื•ืฆืื•ืช. ื”ื•ื ืื™ื ื• ื”ืกืชื‘ืจื•ืช ืœื ื›ื•ื ื•ืช ื”ืชื•ืฆืื” ืื• ืื—ื•ื– ื•ื“ืื•ืช.

ืžืงื•ืจ ื•ืืจื›ื™ื˜ืงื˜ื•ืจื”

ื”ื—ื‘ื™ืœื” ืžื‘ื•ืกืกืช ืขืœ ืžื•ื“ืœ ื”ืื—ื–ื•ืจ Round 2. ืœื ื‘ื•ืฆืข ืื™ืžื•ืŸ ื ื•ืกืฃ ืœืฆื•ืจืš ื”ื™ื™ืฆื•ื.

ื”ืžืงื•ื“ื“ ื›ื•ืœืœ BERT ืขื 8 ืฉื›ื‘ื•ืช, ืจื•ื—ื‘ ื—ื‘ื•ื™ 512, ืฉืžื•ื ื” ืจืืฉื™ ืงืฉื‘, ืจื•ื—ื‘ ื‘ื™ื ื™ื™ื 2048 ื•ืื•ืฆืจ ืžื™ืœื™ื ืฉืœ 32,000 ื˜ื•ืงื ื™ื; ื›ึพ42.26 ืžื™ืœื™ื•ืŸ ืคืจืžื˜ืจื™ื ืœืคื ื™ ื›ื™ืžื•ืช. ื”ื’ืจืฃ ื›ื•ืœืœ ืืช ืฉืจืฉืจืช ื”ืงื™ื“ื•ื“ ื”ืžืœืื”: backbone, ืžืžื•ืฆืข ืœืคื™ ืžืกื›ืช ื”ืงืฉื‘, ื”ื˜ืœื” ื ืœืžื“ืช ืžึพ512 ืœึพ256, LayerNorm ื•ื ืจืžื•ืœ L2.

ืžืงื•ืจ ื”ื™ื™ืฆื•ื ื”ื•ื ืชื™ืงื™ื™ืช final ื‘ื’ืจืกื” 8d7016cf472fb436dbd3ac843ef4d407c28d08f4. ืคืจื˜ื™ ื”ืžืงื•ืจ ื•ื”ื—ืชื™ืžื•ืช ืžื•ืคื™ืขื™ื ื‘ึพmanifest.json.

ื”ืขืจื›ื” ื•ืžื’ื‘ืœื•ืช

ืื™ื›ื•ืช ืื—ื–ื•ืจ ืฉืœ checkpoint ื”ืžืงื•ืจ

ืžื˜ืึพื“ืื˜ืช checkpoint ื”ืžืงื•ืจ ืžื“ื•ื•ื—ืช ืขืœ ืกื˜ ืคื™ืชื•ื— ืฉืœ 410 ืฉืืœื•ืช ื•ึพ12,002 ืžืงื˜ืขื™ื, ืžื“ื“ ื–ื” ื ื‘ื“ืง ืขืœ ืฉืืœื•ืช ืกื™ื ื˜ื˜ื™ื•ืช ื•ืื™ื ื• ืžืฉืงืฃ ืืช ื™ื›ื•ืœื•ืช ื”ืžื•ื“ืœ ืขืœ ืฉืืœื•ืช ืื ื•ืฉื™ื•ืช:

ืžื“ื“ ืขืจืš
nDCG@10 0.4040
MRR 0.3662
Recall@5 0.4683
Recall@10 0.5537

ืืœื” ืžื“ื“ื™ checkpoint ื”ืžืงื•ืจ ื‘ืกื˜ ื”ืคื™ืชื•ื— ืฉืฉื™ืžืฉ ื’ื ืœื‘ื—ื™ืจืชื•. ื”ื ืื™ื ื ื‘ื“ื™ืงื” ืขืฆืžืื™ืช ืฉืœ INT8 ืื• ื”ื•ื›ื—ื” ืœืื™ื›ื•ืช ืขืœ ืžืื’ืจ ื—ื“ืฉ. ื’ื•ื“ืœ ื”ืžืื’ืจ, ื—ืœื•ืงื” ืœืฉื•ืจื•ืช ื‘ืžืงื•ื ืœืžืงื˜ืขื™ื ื•ืกื•ื’ ื”ืฉืืœื•ืช ืขืฉื•ื™ื™ื ืœืฉื ื•ืช ืืช ื”ืชื•ืฆืื•ืช. ืžื•ืžืœืฅ ืœื‘ื“ื•ืง ืฉืืœื•ืช ืžื™ื™ืฆื’ื•ืช ืฉืœ ื”ื™ื™ืฉื•ื ื•ืœื”ืฉื•ื•ืช ื—ื™ืคื•ืฉ ืกืžื ื˜ื™, ืžื™ืœื•ืœื™ ื•ื”ื™ื‘ืจื™ื“ื™.

ืื™ืžื•ืช ื™ื™ืฆื•ื ONNX

ืฉื ื™ ื”ืงื‘ืฆื™ื ืขื‘ืจื• ื‘ื“ื™ืงืช ืชืงื™ื ื•ืช ONNX ื•ื”ืจืฆื” ืขืœ CPU. ื‘ืืจื‘ืขื” ืงืœื˜ื™ ื‘ื“ื™ืงื” ื‘ืขื‘ืจื™ืช ื•ื‘ืืจืžื™ืช, ื‘ืื•ืจื›ื™ื ืฉื•ื ื™ื ืขื“ 256 ื˜ื•ืงื ื™ื, ื”ื•ืฉื•ื• ื”ื•ื•ืงื˜ื•ืจื™ื ืœืžืงื•ื“ื“ PyTorch ื”ืžืงื•ืจื™:

ื™ื™ืฆื•ื ื”ืคืจืฉ ืžื•ื—ืœื˜ ืžืจื‘ื™ ื‘ืงื•ืื•ืจื“ื™ื ื˜ื” ื“ืžื™ื•ืŸ ืงื•ืกื™ื ื•ืก ืžื–ืขืจื™ ืœืžืงื•ืจ
FP32 2.85e-7 0.99999994
INT8 0.01123 0.99863

ื–ื• ื‘ื“ื™ืงืช ืงืจื‘ื” ืžืกืคืจื™ืช ื‘ืืจื‘ืข ื“ื•ื’ืžืื•ืช, ื•ืœื ืžื“ื“ ืื™ื›ื•ืช ืื—ื–ื•ืจ. ื›ื™ืžื•ืช INT8 ืขืฉื•ื™ ืœืฉื ื•ืช ื“ื™ืจื•ื’ื™ื, ื‘ืžื™ื•ื—ื“ ื›ืืฉืจ ื”ืฆื™ื•ื ื™ื ืงืจื•ื‘ื™ื. ืœื ืžืคื•ืจืกื ื›ืืŸ ืžื“ื“ ืžื”ื™ืจื•ืช: ื–ืžื ื™ ืงื™ื“ื•ื“ ื•ื—ื™ืคื•ืฉ ืชืœื•ื™ื™ื ื‘ื—ื•ืžืจื”, ื‘ืื•ืจืš ื”ืงืœื˜ ื•ื‘ืื™ื ื“ืงืก.

ืฉื™ืœื•ื‘ ื‘ื–ื™ืช (Zayit)

ื”ื—ื‘ื™ืœื” ื ื•ืขื“ื” ืœืืคืฉืจ ืฉื™ืœื•ื‘ ื‘ื–ื™ืช. ืœืคื™ ืชื™ืื•ืจ ื”ืชืื™ืžื•ืช ืฉืฉื™ืžืฉ ืœื”ื›ื ืช ื”ื™ื™ืฆื•ื, ืžืžืฉืง v5 ื”ืงื™ื™ื ืžืฆืคื” ืœึพ384 ืžืžื“ื™ื, ืœืžื’ื‘ืœื” ืฉืœ 128 ื˜ื•ืงื ื™ื ื•ืœืงืœื˜ ืœืœื ืงื™ื“ื•ืžื•ืช. ืฉื™ืœื•ื‘ ืžื•ื“ืœ ื–ื” ื“ื•ืจืฉ ืชืžื™ื›ื” ื‘ึพ256 ืžืžื“ื™ื, ื”ื•ืกืคืช ื”ืงื™ื“ื•ืžื•ืช, ืงื™ื“ื•ื“ ืžื—ื“ืฉ ืฉืœ ื”ื•ื•ืงื˜ื•ืจื™ื ื•ื‘ื ื™ื™ืช ืื™ื ื“ืงืก Lucene ื‘ื”ืชืื. ื™ืฉ ืœื‘ื“ื•ืง ื‘ื ืคืจื“ ื ืจืžื•ืœ ื˜ืงืกื˜ ื•ืื™ื›ื•ืช ืื—ื–ื•ืจ ื‘ืจืžืช ืฉื•ืจื”; ืฉื™ื ื•ื™ ืฉื ื”ืงื•ื‘ืฅ ื‘ืœื‘ื“ ืื™ื ื• ืžืฉืœื™ื ืืช ื”ืฉื™ืœื•ื‘.

ื“ื•ื’ืžืื•ืช ื‘ืชืžื•ื ื•ืช

ื”ืชืžื•ื ื•ืช ื”ื‘ืื•ืช ืฆื•ืจืคื• ืœื”ืžื—ืฉืช ื”ืฉื™ืžื•ืฉ. ื”ืŸ ืื™ื ืŸ ืžื“ื“ ื”ืขืจื›ื” ืฉื™ื˜ืชื™ ืื• ื”ืชื—ื™ื™ื‘ื•ืช ืœืชื•ืฆืื•ืช. ืงื™ืฉื•ืจื™ ื”ืžืงื•ืจ.

ื“ื•ื’ืžืช ืฉื™ืžื•ืฉ 1

ื“ื•ื’ืžืช ืฉื™ืžื•ืฉ 2

ื“ื•ื’ืžืช ืฉื™ืžื•ืฉ 3

ื“ื•ื’ืžืช ืฉื™ืžื•ืฉ 4

ื“ื•ื’ืžืช ืฉื™ืžื•ืฉ 5

ื“ื•ื’ืžืช ืฉื™ืžื•ืฉ 6

ื“ื•ื’ืžืช ืฉื™ืžื•ืฉ 7

ื“ื•ื’ืžืช ืฉื™ืžื•ืฉ 8

ืงื•ื‘ืฆื™ ื”ืžืื’ืจ

ืงื•ื‘ืฅ ืชื•ื›ืŸ
seforim-embed-round2-fp32.onnx ื™ื™ืฆื•ื FP32
seforim-embed-round2-int8.onnx ื™ื™ืฆื•ื INT8 ืงื•ืžืคืงื˜ื™ ื‘ื›ื™ืžื•ืช ื“ื™ื ืžื™
tokenizer.json ื”ื˜ื•ืงื ื™ื™ื–ืจ ื”ืชื•ืื ืœืžื•ื“ืœ
manifest.json ื—ื•ื–ื” ื”ืงืœื˜ ื•ื”ืคืœื˜, ืžืงื•ืจ, ืชื•ืฆืื•ืช ืื™ืžื•ืช, ื’ื“ืœื™ื ื•ื—ืชื™ืžื•ืช SHA-256
export_round2_onnx.py ืงื•ื“ ื”ื™ื™ืฆื•ื ื•ื‘ื“ื™ืงื•ืช ื”ืงืจื‘ื” ืœืžืงื•ืจ
LICENSE.md ืชื ืื™ ื”ืจื™ืฉื•ื™ ืœืคื™ ืจื›ื™ื‘

ืžื” ื”ืœืื”?

  • ื”ื•ื›ื—ืช ื”ื™ืชื›ื ื•ืช ืœืžื•ื“ืœ ืฉืคื” ืชื•ืจื ื™ ืงืœ ืžืฉืงืœ.

  • ืื™ืžื•ืŸ ืžื•ื“ืœ ื”ื˜ืžืขื” ื‘ืขืœ ื‘ื™ืฆื•ืขื™ื ืฉื™ืžื•ืฉื™ื™ื.

  • ืฉื™ืคื•ืจ ืื’ืจืกื™ื‘ื™ ืฉืœ ื”ืžื•ื“ืœ ื‘ืืžืฆืขื•ืช ืžืฉื•ื‘ ืื ื•ืฉื™ ืžืžืฉืชืžืฉื™ื.

  • ืคื™ืชื•ื— ืžื•ื“ืœื™ื ื—ื–ืงื™ื ื•ืฉื™ืžื•ืฉื™ื™ื ื™ื•ืชืจ:

    • ื”ืชืžืงื“ื•ืช ื‘ืžื™ืœื™ื ื—ืฉื•ื‘ื•ืช โ€” ืžืชืŸ ืžืฉืงืœ ื’ื‘ื•ื” ื™ื•ืชืจ ืœืžื™ืœื™ื ืื• ืœืžื‘ื ื™ื ืžืฉืžืขื•ืชื™ื™ื, ื›ื’ื•ืŸ ืฆื™ื˜ื•ื˜ื™ื.
    • ื”ืงืฉืจ ืืจื•ืš โ€” ืชืžื™ื›ื” ื‘ึพ4K ื˜ื•ืงื ื™ื.
    • ืื•ืจืš ื•ืงื˜ื•ืจ ื’ืžื™ืฉ โ€” ืืคืฉืจื•ืช ืœื”ืฉืชืžืฉ ืจืง ื‘ึพ32/64/128/256/512/1024 ื”ืขืจื›ื™ื ื”ืจืืฉื•ื ื™ื ืฉืœ ื”ื•ื•ืงื˜ื•ืจ, ืชื•ืš ืจื™ื›ื•ื– ื”ืžื™ื“ืข ื”ื—ืฉื•ื‘ ื‘ืชื—ื™ืœืช ื”ื•ื•ืงื˜ื•ืจ.
    • ื”ืฉื•ื•ืื” ื™ืฉื™ืจื” ื‘ื™ืŸ ืฉืื™ืœืชื” ืœืงื˜ืข โ€” ืชืžื™ื›ื” ื‘ื”ื–ื ืช ืฉืื™ืœืชื” ื•ืงื˜ืข ื™ื—ื“ ืœืงื‘ืœืช ืฆื™ื•ืŸ ืจืœื•ื•ื ื˜ื™ื•ืช ืžื“ื•ื™ืง ื™ื•ืชืจ.
  • ืื™ืžื•ืŸ ืžืฉืคื—ืช ืžื•ื“ืœื™ื ื‘ื’ื“ืœื™ื ืฉื•ื ื™ื, ื”ืžืฉืชืคื™ื ืžืจื—ื‘ ืกืžื ื˜ื™ ืžืฉื•ืชืฃ ื•ืžืืคืฉืจื™ื ื”ื—ืœืคื” ืžืœืื” ื‘ื™ื ื™ื”ื:

    • Sinai โ€” ืžื•ื“ืœ ืžื”ื™ืจ ื•ืงืœ ืžืฉืงืœ [ืžืฉื•ืขืจ: ~30M ืคืจืžื˜ืจื™ื].
    • OkerHarim โ€” ืžื•ื“ืœ ื—ื–ืง ื•ืžื“ื•ื™ืง [ืžืฉื•ืขืจ: ~80M ืคืจืžื˜ืจื™ื]. ื ื™ืชืŸ ืœื™ืฆื•ืจ ืื™ื ื“ืงืก ื‘ืืžืฆืขื•ืช OkerHarim ื•ืœื‘ืฆืข ื—ื™ืคื•ืฉ ื‘ืืžืฆืขื•ืช Sinai, ืื• ืœื”ืคืš, ืœืœื ืฆื•ืจืš ื‘ื™ืฆื™ืจืช ืื™ื ื“ืงืก ืžื—ื“ืฉ, ืžืฉื•ื ืฉืฉื ื™ ื”ืžื•ื“ืœื™ื ืžืžืคื™ื ื˜ืงืกื˜ื™ื ืœืื•ืชื• ืžืจื—ื‘ ืกืžื ื˜ื™.

ืจื™ืฉื•ื™ ื•ื™ื™ื—ื•ืก

ื›ืœ ืฉื™ืžื•ืฉ ืžืกื—ืจื™ ืืกื•ืจ ืœืœื ืื™ืฉื•ืจ ืžืคื•ืจืฉ ืžืจืืฉ ื•ื‘ื›ืชื‘ ืžืืจื™ืืœ ื“ื ื™ืืœื™ โ€” arieldaniely@gmail.com. ืืกื•ืจ ืœืฉืœื‘ ืื• ืœื”ืฉืชืžืฉ ื‘ืžื•ื“ืœ ื‘ื›ืœ ืฆื•ืจื” ื‘ืžืขืจื›ืช, ืชื•ื›ื ื” ืื• ืฉื™ืจื•ืช ืœืžื˜ืจื•ืช ืจื•ื•ื—, ื’ื ืื ื”ืจื›ื™ื‘ ืฉื‘ื• ื”ื•ื ืžื•ื˜ืžืข ื—ื™ื ืžื™. ืจืื• ื”ืจื™ืฉื™ื•ืŸ ื”ืžืœื.

ื”ืžื•ื“ืœ ื•ื”ื˜ื•ืงื ื™ื™ื–ืจ ืžื•ืคืฆื™ื ื‘ื’ืจืกื” ื–ื• ืชื—ืช Meivin Non-Commercial License 1.0. ื”ืื™ืกื•ืจ ื›ื•ืœืœ ืฉื™ืœื•ื‘ ื™ืฉื™ืจ ืื• ืขืงื™ืฃ, API, ืชื•ืกืฃ, ืฉื™ืจื•ืช ืžืจื•ื—ืง ื•ืฉื™ืžื•ืฉ ื‘ื•ื•ืงื˜ื•ืจื™ื ืื• ื‘ืื™ื ื“ืงืกื™ื ื›ืจื›ื™ื‘ ื‘ืžืขืจื›ืช ืžืกื—ืจื™ืช. ืื™ืฉื•ืจ ื”ื•ืจื“ื” ืื• ื’ื™ืฉื” ืœืžืื’ืจ ืื™ื ื• ืื™ืฉื•ืจ ืœืฉื™ืžื•ืฉ ืžืกื—ืจื™. ื”ืคืฆื” ื•ืขื™ื‘ื•ื“ื™ื ืžื—ื™ื™ื‘ื™ื ืฉืžื™ืจืช ื”ืจื™ืฉื™ื•ืŸ ื•ื”ื™ื™ื—ื•ืก. ืงื•ื“ ืžืงื•ืจื™ ื ืœื•ื•ื” ื ืฉืืจ ื›ืคื•ืฃ ืœึพPersonal Use License 1.0 ืฉื‘ืงื•ื‘ืฅ ื”ืจื™ืฉื™ื•ืŸ; ื–ื›ื•ื™ื•ืช ืฆื“ ืฉืœื™ืฉื™ ื ืฉืžืจื•ืช.

ื™ื™ื—ื•ืก: ืืจื™ืืœ ื“ื ื™ืืœื™ โ€” Judaic Semantic Embedding project, Otzaria. ื”ื™ื™ื—ื•ืก ืžืชื™ื™ื—ืก ืœืคื™ืชื•ื— ื”ืืจื˜ื™ืคืงื˜ื™ื ื•ืื™ื ื• ื˜ืขื ื” ืœืžื—ื‘ืจื•ึผืช ืฉืœ ื˜ืงืกื˜ื™ ื”ืžืงื•ืจ.

ืชื™ืขื•ื“ ื”ืื™ืžื•ืŸ

ื“ื•ื— ื”ืื™ืžื•ืŸ ื”ืžืคื•ืจื˜ ืžืชืขื“ ืืช ืžืกืœื•ืœ ื”ืžืฉืงื•ืœื•ืช ืฉืœ Round 2, ืื™ืžื•ื ื™ ื”ึพMLM, ื”ื™ื™ืฉื•ืจ ื”ืขื‘ืจื™โ€“ืืจืžื™, ืกื‘ื‘ื™ ื”ืื—ื–ื•ืจ ื•ื”ื ื™ืกื•ื™ื™ื ืฉืœื ืงื•ื“ืžื•. ื”ื•ื ื›ื•ืœืœ ื”ื™ืคืจึพืคืจืžื˜ืจื™ื, ื’ืจืคื™ื, ื ืชื•ื ื™ ืžื“ื™ื“ื” ื•ืžื’ื‘ืœื•ืช ื”ื”ืขืจื›ื”.

MLM fixed evaluation

ื”ื’ืจืฃ ืžืฆื™ื’ ื‘ื“ื™ืงืช MLM ืงื‘ื•ืขื” ื‘ืื•ืจื›ื™ื ืžืขื•ืจื‘ื™ื; ื”ื•ื ืื™ื ื• ืžื“ื“ ื”ืฆืœื—ื” ืฉืœ ื—ื™ืคื•ืฉ ืžืงื•ืจื•ืช. ืชื•ืฆืื•ืช ื”ืื—ื–ื•ืจ ื•ืคืจื•ื˜ื•ืงื•ืœื™ ื”ื‘ื“ื™ืงื” ืžื•ืคื™ืขื™ื ื‘ื“ื•ื—.

Languages and textual coverage | ืฉืคื•ืช ื•ื ื™ื‘ื™ ืœืฉื•ืŸ

Metadata Coverage Evidence and interpretation
he โ€” Hebrew Biblical, Mishnaic/Rabbinic, medieval and modern Hebrew Tanakh, Mishnah, halakhic and commentarial literature, and modern HeQ text
tmr โ€” Jewish Babylonian Aramaic (ca. 200โ€“1200 CE) Aramaic in the Babylonian Talmud Bavli corpus bucket; Bavli also contains Hebrew
jpa โ€” Jewish Palestinian Aramaic Palestinian rabbinic Aramaic represented in the Jerusalem Talmud Yerushalmi corpus bucket; Yerushalmi also contains Hebrew
arc โ€” Official/Imperial Aramaic code Biblical Aramaic within Daniel and Ezra; broad Aramaic discovery label Both books were verified in retained training rows; not every verse is Aramaic

ื”ืกื™ืžื•ืŸ ื”ื™ื“ื ื™ he, tmr, jpa, arc ืžืชืื™ื ืœื”ื™ืงืฃ ื”ืžืงื•ืจื•ืช. tmr ื”ื•ื ื”ืืจืžื™ืช ื”ื™ื”ื•ื“ื™ืช ื”ื‘ื‘ืœื™ืช, jpa ื”ืืจืžื™ืช ื”ื™ื”ื•ื“ื™ืช ื”ืืจืฅึพื™ืฉืจืืœื™ืช, ื•ึพarc ืžืฉืžืฉ ื›ืืŸ ื’ื ืœืกื™ืžื•ืŸ ื”ืืจืžื™ืช ื”ืžืงืจืื™ืช. ื“ื ื™ืืœ ื•ืขื–ืจื ื›ื•ืœืœื™ื ืขื‘ืจื™ืช ื•ืืจืžื™ืช; ื‘ื‘ืœื™ ื•ื™ืจื•ืฉืœืžื™ ื”ื ืฉืžื•ืช ืงื•ืจืคื•ืกื™ื, ืœื ืฉืžื•ืช ืฉืคื•ืช. ืื™ืŸ ืœืกืžืŸ ืืช ื›ืœ ื”ืชืœืžื•ื“ื™ื ื›ืฉืคื” ืื—ืช ืื• ืืช ื›ืœ ืกืคืจ ื“ื ื™ืืœ ื›ืืจืžื™ืช.

The arc ISO label is historical and is not a precise label for every Aramaic variety in the corpus. Targumic and Zoharic/literary Aramaic are also represented in the broader Judaic source collection; they are described here rather than assigned unrelated modern Aramaic codes. Presence of texts establishes training coverage, not independently benchmarked mastery of each dialect. No per-language quality scores or exact percentages of Aramaic are available, because corpus buckets are not language annotations. English instructions in this repository do not imply English model support.

See aggregate corpus checks. Language code references: Jewish Babylonian Aramaic, Jewish Palestinian Aramaic, and Biblical Hebrew/Aramaic language codes in BHSA.

Downloads and parameter display | ื”ื•ืจื“ื•ืช ื•ืคืจืžื˜ืจื™ื

The full FP32 checkpoint is distributed as root model.safetensors, enabling Hugging Face to inspect its tensor headers. Hugging Face currently reports 42,260,992 parameters in its native Safetensors display. The direct PyTorch count is 42,260,993, including the scalar training parameter logit_scale. The one-parameter difference does not affect embeddings or model size in practical terms. Both counts refer to the same released file.

The count includes stored training parameters; ONNX can omit unused components. INT8 is a numerical export of the same model, not a reduction in architectural parameter count.

Use the documented loader/configuration together with the weights. The Hub counts server-side requests to designated query files such as config.json; it does not count unique users or every arbitrary file download. ONNX integrations below load the real configuration as part of model initialization. Direct downloads of ONNX files alone may not be represented in this counter. Historical missed requests cannot be reconstructed from repository metadata.

The model page shows recent downloads; Hub API also exposes downloadsAllTime for cumulative downloads, when available. The release includes an example that reads both fields live, without hardcoding a launch-day number. See Hugging Face counting rules.

Citation

@misc{daniely2026meivinembed,
 author = {Ariel Daniely},
 title = {Meivin Embed 0.1: Semantic Retrieval for Hebrew and Aramaic Judaic Texts},
 year = {2026},
 url = {https://huggingface.co/ArieLLL123/Meivin-Embed-0.1}
}
Downloads last month
3
Safetensors
Model size
42.3M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ArieLLL123/Meivin-Embed-0.1

Collection including ArieLLL123/Meivin-Embed-0.1