File size: 3,138 Bytes
908f811
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
# Match-vector based hard negative policy

This file defines how fact-retrieval training should consume the enriched InfoSeek fact maps.

## 1. Source fields

For each InfoSeek numeric/time-style training row:

```text
positive = selected_positive_fact_ids
negative metadata = hard_negative_meta_v2[fact_id]
```

Do not rely on the old negative type label alone. The type label is useful as provenance/debug metadata, but the training sampler should use `match_vector`.

## 2. Match vector

Each hard negative fact has:

```json
{
  "entity": true/false,
  "property": true/false,
  "qualifier": true/false/null,
  "value": true/false/null
}
```

Order shorthand:

```text
entity / property / qualifier / value
```

Examples:

```text
11n0 = same entity, same property, no qualifier constraint, wrong value
1110 = same entity, same property, same qualifier, wrong value
10n1 = same entity, different property, same value
01n0 = different entity, same property, wrong value
```

`null` means the dimension is not applicable or cannot be judged. It is not the same as `false`.

## 3. Positive-like negatives

The following vectors must not be used as negatives:

```text
1111
11n1
```

These mean entity/property/value all match the selected positive, so they are likely duplicate positives or unresolved aliases.

Current sanity after vector recomputation:

```text
val:   missing_vector = 0, positive_like_negative = 0

train: missing_vector = 0, positive_like_negative = 0
```

## 4. Training negative priority

Recommended sampling priority:

### Tier A: strongest hard negatives

Same entity and same property, but wrong value or wrong qualifier:

```text
1110

11n0

1100

1101
```

These directly teach the fact retriever not to select a fact just because the entity and relation are right.

### Tier B: same entity, different property

Same subject entity, but the requested property is different:

```text
10n1
10n0
10nn
1001
1010

1011

```

These are useful because the visual entity is correct, but the retrieved fact answers the wrong question.

### Tier C: same property, different entity

Same or similar property, but from another subject:

```text
01n0
0100
0110
01n1
0101
0111
```

These teach the model not to retrieve a globally plausible property fact from the wrong entity.

### Tier D: mixed/easy negatives

Everything else:

```text
00n0
00n1
0010
0011
```

Use only as filler if there are not enough Tier A/B/C negatives.

## 5. Dataloader rule

For each row:

```text
positive_ids = selected_positive_fact_ids

negative_ids =
  sample Tier A first
  then Tier B
  then Tier C
  then Tier D if needed
```

The sampler should keep the vector metadata, because later codebook analysis will compute:

```text
same_code_prefix_len(positive_fact, negative_fact)
```

This will be added after fact embeddings/codebooks are built.

## 6. Important caveat

The existing type labels may still contain noisy names such as `same_entity_property_matched_wrong_value` even when the final `match_vector.property` is `false`.

That is expected after the correction. Training must follow `match_vector`, not the legacy label.