1
Token序列:  [BOS] [产] [品] [介] [绍] [IMG] [IMG] ... [EOS] [PAD] [PAD]
2
attention:    1    1    1    1    1    1     1   ...   1     0     0
3
                                                        ↑
4
                                               提取这个位置的向量 [D]
5

6
为什么用 EOS?
7
├─ EOS token 通过因果注意力"看过"整个序列的所有信息
8
└─ 是对整个输入的最佳压缩表示

1
def _pooling_last(hidden_state, attention_mask):
2
    # 翻转 mask，找到最后一个 1 的位置 (EOS 位置)
3
    flipped = attention_mask.flip(dims=[1])
4
    last_pos = attention_mask.shape[1] - flipped.argmax(dim=1) - 1
5
    row = torch.arange(hidden_state.shape[0])
6
    return hidden_state[row, last_pos]  # 取出 EOS 向量

1
# 生成查询向量
2
query = {"text": "如何配置网络设置？", "instruction": "检索产品文档"}
3
query_embedding = embedder.process([query], normalize=True)
4
# 输出: [1, 2048] 的归一化向量

1
# 构造的 prompt 格式
2
{
3
    "role": "system",
4
    "content": "Judge whether the Document meets the requirements based on the
5
                Query and the Instruct. Note that the answer can only be 'yes' or 'no'."
6
},
7
{
8
    "role": "user",
9
    "content": [
10
        "<Instruct>: 检索相关产品文档",
11
        "<Query>: 如何配置网络设置？",
12
        "<Document>: 网络配置步骤：1. 打开设置面板..."
13
    ]
14
}

1
def get_binary_linear(self, model, token_yes, token_no):
2
    lm_head_weights = model.lm_head.weight.data
3
    weight_yes = lm_head_weights[token_yes]  # "yes" token 的权重
4
    weight_no = lm_head_weights[token_no]    # "no" token 的权重
5

6
    linear_layer = torch.nn.Linear(D, 1, bias=False)
7
    linear_layer.weight[0] = weight_yes - weight_no  # 差值作为分类权重
8
    return linear_layer
9

10
def compute_scores(self, inputs):
11
    hidden_state = self.model(**inputs).last_hidden_state[:, -1]  # 取最后 token
12
    scores = self.score_linear(hidden_state)  # 线性变换
13
    scores = torch.sigmoid(scores)  # 归一化到 [0, 1]
14
    return scores