Qwen2.5-0.5B dense retriever (bi-encoder, bidirectional attention, mean pooling), trained on synthetic ClueWeb22 queries produced by a query generator aligned via Direct Preference Optimization (DPO) using a point-wise re-ranker (bge-reranker-v2-m3) reward.

Base model: jmvcoelho/Qwen2.5-0.5B-bidirectional-attn-mntp.

Results (MS MARCO, 1M training pairs)

MARCO Doc MRR@100	MARCO Passage nDCG@10	TREC-DL19 nDCG@10	TREC-DL20 nDCG@10
0.3727	0.3286	0.6171	0.5337

Usage

python

import torch, torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer

repo = "jmvcoelho/Qwen2.5-0.5B-retriever-ranker-dpo-1M"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModel.from_pretrained(repo, trust_remote_code=True).eval()

def encode(texts):
    batch = tok(texts, padding=True, truncation=True, max_length=512, return_tensors="pt")
    with torch.no_grad():
        out = model(**batch).last_hidden_state
    mask = batch["attention_mask"].unsqueeze(-1).float()
    emb = (out * mask).sum(1) / mask.sum(1)      # mean pooling
    return F.normalize(emb, dim=-1)              # L2 normalize; score = cosine / dot

q = encode(["what is dense retrieval"])
d = encode(["Dense retrieval uses dual encoders ..."])
print((q @ d.T).item())