README.md

---
license: apache-2.0
library_name: transformers
pipeline_tag: text-classification
tags:
  - reranker
  - cross-encoder
  - dual-passage-classifier
  - multi-hop-qa
  - pytorch
datasets:
  - musique
model-index:
  - name: Dual Passage Classifier
    results:
      - task:
          type: text-classification
          name: Passage Reranking (Multi-hop QA)
        dataset:
          name: MuSiQue-Full
          type: musique
          split: validation
        metrics:
          - type: MAP
            value: TODO
            name: Mean Average Precision
---

# Dual Passage Classifier (DPC) for Multi-hop QA 🔎📑

**Dual Passage Classifier (DPC)** 是一個 *cross-encoder* reranker，  
輸入 **(Question, Passage 1, Passage 2)**，輸出一維分數，判斷這對段落對回答該問題的「共同貢獻程度」：

| label | 定義 |  Margin label |
|-------|------|--------------|
| positive | d1 & d2 都必要 | 0 |
| neutral  | 其中一段必要 |1 |
| negative | 都不重要     |2 |

本模型將 **三段式 MarginRankingLoss** 混合，  
並在 MuSiQue 上把 baseline（`naver/trecdl22-crossencoder-debertav3`）的模型。
這是訓練在 Musique Dataset 的模型
---

## Quick-start

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

tokenizer = AutoTokenizer.from_pretrained("QQhahaha/musique-deberta-v3-large-MLP-2-Marginloss-ratio7")
model     = AutoModel.from_pretrained("QQhahaha/musique-deberta-v3-large-MLP-2-Marginloss-ratio7").eval()

def score(q, d1, d2):
    text = f"{q} [SEP] {d1} [SEP] {d2}"
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
    with torch.no_grad():
        logit = model(**inputs).last_hidden_state[:,0,:] @ model.classifier[0].weight.T
    return torch.sigmoid(logit).item()         # 0~1，越高代表正向