File size: 7,595 Bytes
b678667
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
---
license: cc-by-sa-4.0
language:
- he
pretty_name: Abstractive QA for LLM Evaluation
task_categories:
- question-answering
- summarization
size_categories:
- n<1K
tags:
- hebrew
- grounding
- hallucination-detection
- attribution
- faithfulness
---

# Abstractive QA for LLM evaluation

A Hebrew-language dataset of 132 question-answer records, built for testing
whether a language model's generated answer is grounded in its source
document rather than hallucinated. Each record has a `question`, a
`source_text`, and a `reference_answer` whose individual claims are each
backed by exact quoted spans from `source_text`. Use it to check your own
model the same way: generate an answer, break it into claims, and see
whether each claim traces back to real text in the source.

## Contents

- [Quick start](#quick-start)
- [Support sentences: how grounding works](#support-sentences-how-grounding-works)
- [Field reference](#field-reference)
- [Key numbers](#key-numbers)
- [Annotator agreement](#annotator-agreement)
- [License](#license)
- [Credits](#credits)

## Quick start

```python
import json

with open("abstractive_qa_for_llm_eval.json", encoding="utf-8") as f:
    records = json.load(f)

record = records[0]
print("Question:", record["question"])
print("Reference answer:", record["reference_answer"])
```

The file is a single JSON array of 132 records. Load it once and index
into it like any list.

## Support sentences: how grounding works

`reference_answer` is not graded as one block of text. It is split into
`claims`, one sentence-level statement each. Every claim carries an
`attribution` array: one or more exact spans copied from `source_text`
that support it. Each span gives a start and end character offset into
`source_text`, so you can slice the source string and get the exact
quoted text back.

Here is a real claim from the dataset, with two support sentences from
two different parts of the source document:

```
claim: החשש היה שהאיסור על קיום כיתות נפרדות על רקע "זהות מגדרית" יפורש
       כאיסור על הפרדה בין בנים לבנות.

attribution[0]: start_char=1326, end_char=1603, verified=true
  source_text[1326:1603] == <מיכל חורין: אני פשוט רוצה לחדד: הנוסח היום
  מדבר טעמים של נטייה מינית או זהות מגדרית, בגלל שאני לא מכירה את המינוח
  "זהות מגדרית" אני רק רוצה לוודא שלא מדובר מטעמים של מין בעצם, שזהות
  מגדרית זה משהו אחר, על מנת שיתאפשר לנו להמשיך להפריד בנים ובנות במוסדות
  של החינוך הממלכתי-דתי.

attribution[1]: start_char=4068, end_char=4565, verified=true
  source_text[4068:4565] == " כפי שנאמר כאן זה משתלב ברישה שאומרת: "5
  (א) רשות חינוך מקומית, מוסד חינוך או אדם הפועל מטעמם, לא יפלו תלמיד
  מטעמים עדתיים, מטעמים של ארץ מוצא מטעמים ...
```

`source_text[start_char:end_char]` equals `source_excerpt` exactly. That
is what `verified: true` means: the span was checked to match the source
text character for character.

To score your own model, do the same two steps on its output: split the
generated answer into individual claims, then search for a matching span
in the source text for each claim. A claim you cannot find in the source
is likely a hallucination. A claim you can find and quote is grounded.

## Field reference

### Top-level fields (one entry per record)

| Field | Type | Meaning |
| --- | --- | --- |
| `id` | string | Short unique ID for the record. |
| `source_task_id` | string | ID of the underlying annotation task this record came from. |
| `genre` | string | Text genre: `dialogue`, `encyclopedic`, or `journalistic`. |
| `difficulty` | integer | Question difficulty: 0 = retrieval, 1 = simple, 2 = complex. |
| `document_id` | string | ID of the source document. |
| `question` | string | The question, in Hebrew. |
| `source_text` | string | The full source document the question is about. |
| `reference_answer` | string | The ground-truth answer, in Hebrew. |
| `claims` | array | The reference answer, split into individually attributed claims. See below. |
| `human_ratings` | array | Quality scores from four human annotators. See below. |

### `claims[]`

| Field | Type | Meaning |
| --- | --- | --- |
| `claim_id` | integer | Position of this claim within the record, starting at 0. |
| `text` | string | The claim text, a sentence-level piece of the reference answer. |
| `attribution` | array | One or more support spans backing this claim. See below. |

### `attribution[]`

| Field | Type | Meaning |
| --- | --- | --- |
| `sentence_id` | integer | ID of the sentence in `source_text` that this span comes from. |
| `source_excerpt` | string | The quoted text, copied exactly from `source_text`. |
| `start_char` | integer | Start offset of the span in `source_text`. |
| `end_char` | integer | End offset of the span in `source_text` (exclusive). |
| `verified` | boolean | Whether `source_text[start_char:end_char]` was checked to equal `source_excerpt`. |
| `verification_method` | string | How the check was done, for example `auto`. |

### `human_ratings[]`

| Field | Type | Meaning |
| --- | --- | --- |
| `annotator_id` | string | ID of the annotator who gave this rating, for example `annotator_1`. |
| `scores` | object | Five scores, each 0-2. Keys: `coherence`, `relevance`, `completeness`, `faithfulness`, `attribution`. |

Each score is 0, 1, or 2: 0 means incorrect or unsupported, 1 means
partially correct, 2 means fully correct.

## Key numbers

Counted directly from the file:

| Metric | Value |
| --- | --- |
| Records | 132 |
| Claims | 244 |
| Attribution spans | 286 |
| Attribution spans verified | 286 of 286 (100%) |
| Human rating entries | 528 (4 annotators x 132 records) |

Genre distribution:

| Genre | Records |
| --- | --- |
| dialogue | 47 |
| encyclopedic | 44 |
| journalistic | 41 |

Difficulty distribution:

| Difficulty | Meaning | Records |
| --- | --- | --- |
| 0 | retrieval | 39 |
| 1 | simple | 45 |
| 2 | complex | 48 |

## Annotator agreement

Two numbers describe how consistently the four annotators rated the same
records. They are given together because they tell different stories.

- **Perfect agreement: 0.379** (50 of 132 records). This is the fraction
  of records where all four annotators gave the exact same score, on all
  five dimensions. It is easy to understand, but it does not correct for
  chance. Most scores in this dataset are the maximum value, so
  annotators can match by both giving the top score without truly
  agreeing on quality.
- **Krippendorff's alpha (ordinal): 0.348 pooled across all dimensions.**
  By dimension: coherence 0.057, relevance 0.373, completeness 0.477,
  faithfulness 0.375, attribution 0.311. Alpha corrects for chance
  agreement, and is the standard measure for rating reliability. A
  published guideline treats alpha >= 0.800 as reliable and 0.667-0.800
  as tentative; below 0.667 is unreliable. Every dimension here falls
  below 0.667.

In short: annotators looked consistent by the simple measure, but the
chance-corrected measure shows weak agreement across every rating
dimension. Anyone using `human_ratings` should keep this in mind and not
treat the scores as a precise ground truth.

## License

This dataset is licensed under CC BY-SA 4.0. See [LICENSE.md](LICENSE.md).

## Credits