Views
No views yet
ℹ️ Bu, Werea-co/DUSUNEN-Rota-270M-v3 modelinin GoktugD profilindeki aynasıdır; asıl geliştirme Werea-co altında sürer.
mteb/TurHistQuadRetrieval
(1.024 sorgu / 1.213 belge), aynı değerlendirici, bf16:| Model | Params | nDCG@10 | MRR@10 | Recall@10 | Recall@100 |
|---|---|---|---|---|---|
| multilingual-e5-base | 278M | 0.499 | 0.697 | 0.531 | 0.728 |
| DUSUNEN Rota 270M v3 | 270M | 0.454 | 0.662 | 0.470 | 0.573 |
| Harrier OSS 270M (taban) | 270M | 0.434 | 0.631 | 0.459 | 0.535 |
1import torch, torch.nn.functional as F
2from transformers import AutoModel, AutoTokenizer
3
4tok = AutoTokenizer.from_pretrained("Werea-co/DUSUNEN-Rota-270M-v3")
5model = AutoModel.from_pretrained("Werea-co/DUSUNEN-Rota-270M-v3", torch_dtype=torch.bfloat16).cuda().eval()
6
7def embed(texts, prefix):
8 e = tok([prefix+t for t in texts], padding=True, truncation=True, max_length=192, return_tensors="pt").to("cuda")
9 with torch.inference_mode():
10 h = model(**e).last_hidden_state
11 m = e["attention_mask"].unsqueeze(-1).float()
12 return F.normalize((h.float()*m).sum(1)/m.sum(1), dim=-1)
13
14q = embed(["Türkiye'nin başkenti neresi?"], "query: ")
15d = embed(["Ankara, Türkiye'nin başkentidir."], "passage: ")
16print((q@d.T).item())