TR Embedding FT: BGE-M3, jina-v3, nomic-embed TR Adaptation + MTEB-TR Eval

# === BGE-M3 TR Fine-Tuning ===
from sentence_transformers import SentenceTransformer, InputExample, losses
from torch.utils.data import DataLoader
 
model = SentenceTransformer("BAAI/bge-m3", device="cuda")
 
# Dataset: (query, positive_doc, negative_doc) triplet
train_examples = []
for query, pos_doc, neg_docs in tr_dataset:
    for neg in neg_docs[:7]:    # 1 pos + 7 hard negatives
        train_examples.append(InputExample(texts=[query, pos_doc, neg]))
 
train_dataloader = DataLoader(train_examples, batch_size=8, shuffle=True)
 
# Loss — MultipleNegativesRankingLoss (InfoNCE variant)
loss = losses.MultipleNegativesRankingLoss(model)
 
model.fit(
    train_objectives=[(train_dataloader, loss)],
    epochs=3,
    warmup_steps=500,
    optimizer_params={"lr": 2e-5},
    output_path="bge-m3-tr-finetuned",
)
 
# Eval — MTEB-TR
from mteb import MTEB
benchmark = MTEB(tasks=["mteb_tr_*"])
results = benchmark.run(model)