oridror/bge-m3-hebrew-r1-myd-r1

bge-m3-hebrew-r1-myd-r1

MYD Round-1 Hebrew fine-tune of BAAI/bge-m3 on 15,142 persona↔CEO Q-A pairs extracted from the MYD synthetic dialog panel (CEOs: Rafael, Adel, Antonio).

Usage

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("oridror/bge-m3-hebrew-r1-myd-r1")
queries  = ["מה זה MedBed?"]
passages = ["MedBed הוא פרויקט לריפוי הוליסטי..."]
q_emb = model.encode(queries, normalize_embeddings=True)
p_emb = model.encode(passages, normalize_embeddings=True)
sim = (q_emb @ p_emb.T)[0][0]
print(sim)

Important: This model (BGE-M3) does not require query/passage prefixes. Encode text directly.

Eval (n=500 held-out Hebrew Q-A pairs)

Training

• —Base: BAAI/bge-m3
• —Loss: MultipleNegativesRankingLoss (in-batch negatives, scale=20)
• —Epochs: 2
• —LR: 2e-5 with 100 warmup steps
• —Hardware: NVIDIA A100 80GB PCIe (RunPod)
• —Run: myd-r1-runpod-5-models (2026-04-22)

Data

Extracted from 6-ai/synthetic-panel/output/dialogs/all_dialogs.jsonl — 3,925 generated Hebrew dialogs (persona + CEO turns). Every persona-role turn paired with the immediately-following CEO-role turn yielded 15,642 Q-A pairs. Split: 15,142 train / 500 eval (seed 42).

License

Apache-2.0 (inherited from base model).

Part of MYD

This model is part of the MYD generative-system stack. Routed via myd-router policy as embed.he candidate.