aryopg commited on
Commit
76fc567
·
verified ·
1 Parent(s): 05eb03f

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +104 -0
README.md CHANGED
@@ -1,4 +1,17 @@
1
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  dataset_info:
3
  features:
4
  - name: question
@@ -21,3 +34,94 @@ configs:
21
  - split: train
22
  path: data/train-*
23
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: mit
3
+ language:
4
+ - en
5
+ task_categories:
6
+ - question-answering
7
+ pretty_name: Parametric & Arithmetic Eval
8
+ size_categories:
9
+ - n<1K
10
+ tags:
11
+ - retrieval-heads
12
+ - mechanistic-interpretability
13
+ - parametric-knowledge
14
+ - ablation
15
  dataset_info:
16
  features:
17
  - name: question
 
34
  - split: train
35
  path: data/train-*
36
  ---
37
+
38
+ # Parametric & Arithmetic Eval
39
+
40
+ A 600-example control set for testing whether ablated attention heads are
41
+ **retrieval-specific** rather than generically important for model output. Every
42
+ question is answerable from the model's own parametric knowledge or by direct
43
+ computation — none require retrieving information from an in-context document.
44
+
45
+ This dataset accompanies **LOCOS** (Logit-Contribution Scoring). A retrieval-head
46
+ detector is only meaningful if ablating the heads it identifies degrades
47
+ *contextual retrieval* while leaving *non-retrieval* abilities intact. We measure
48
+ that gap with the **Dissociation Score** at ablation depth $k$:
49
+
50
+ $$\mathrm{DS}(k) = \Delta R(k) - \Delta P(k)$$
51
+
52
+ where $\Delta R(k) = (R_0 - R(k))/R_0$ is the relative drop in NoLiMa ROUGE-L
53
+ (retrieval), and $\Delta P(k) = (P_0 - P(k))/P_0$ is the relative drop in
54
+ aggregate parametric accuracy (mean over the three sources below). $R_0, P_0$ are
55
+ the unablated baselines. A high DS means the heads are retrieval-specific:
56
+ ablating them hurts retrieval far more than parametric recall or arithmetic.
57
+
58
+ ## Composition
59
+
60
+ 600 examples, 200 per source:
61
+
62
+ | `source` | Count | Task | Example question | Example answer |
63
+ |----------------|-------|-----------------------------|-----------------------------------------------|----------------|
64
+ | `city_country` | 200 | Parametric factual recall | *Which country does Tokyo belong to?* | `Japan` |
65
+ | `popqa` | 200 | Parametric factual recall | *Who is the father of Brandon Lee?* | `Bruce Lee` |
66
+ | `arithmetic` | 200 | Two-operand add/subtract | *What is (5 + 6) - 3?* | `8` |
67
+
68
+ ## Fields
69
+
70
+ | Field | Type | Description |
71
+ |------------|--------|-------------------------------------------------------------------|
72
+ | `question` | string | The prompt question. |
73
+ | `answer` | string | Reference answer (short form). |
74
+ | `source` | string | One of `city_country`, `popqa`, `arithmetic`. |
75
+ | `index` | int64 | Global row index (0–199 city_country, 200–399 popqa, 400–599 arithmetic). |
76
+
77
+ ## Provenance
78
+
79
+ - **`city_country`** — top-200 cities by population from
80
+ [`WorkWithData/cities`](https://huggingface.co/datasets/WorkWithData/cities),
81
+ reformatted as country-identification questions (country names normalised to
82
+ common English forms).
83
+ - **`popqa`** — most-popular entries from
84
+ [`akariasai/PopQA`](https://huggingface.co/datasets/akariasai/PopQA), ranked by
85
+ the geometric mean of subject/object popularity and deduplicated by
86
+ subject–object pair.
87
+ - **`arithmetic`** — random sample (seed 42) from
88
+ [`EleutherAI/arithmetic`](https://huggingface.co/datasets/EleutherAI/arithmetic)
89
+ (`arithmetic_1dc`), with the `Question:`/`Answer:` scaffolding stripped.
90
+
91
+ ## Usage
92
+
93
+ ```python
94
+ from datasets import load_dataset
95
+
96
+ ds = load_dataset("aryopg/parametric-arithmetic-eval", split="train")
97
+ city_country = ds.filter(lambda r: r["source"] == "city_country")
98
+ ```
99
+
100
+ ## License
101
+
102
+ Released under the **MIT License**. The dataset is derived from openly available
103
+ sources; please also respect the licenses of the upstream datasets linked above.
104
+
105
+ ## Citation
106
+
107
+ If you use this dataset, please cite LOCOS and PopQA:
108
+
109
+ ```bibtex
110
+ @misc{gema2026locos,
111
+ title = {Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads},
112
+ author = {Gema, Aryo Pradipta and Alex, Beatrice and Minervini, Pasquale},
113
+ year = {2026},
114
+ note = {Preprint forthcoming}
115
+ }
116
+ ```
117
+
118
+ ```bibtex
119
+ @article{mallen2023llm_memorization,
120
+ title = {When Not to Trust Language Models: Investigating Effectiveness and Limitations of Parametric and Non-Parametric Memories},
121
+ author = {Mallen, Alex and Asai, Akari and Zhong, Victor and Das, Rajarshi and Hajishirzi, Hannaneh and Khashabi, Daniel},
122
+ journal = {arXiv preprint},
123
+ year = {2022}
124
+ }
125
+ ```
126
+
127
+ > **Note:** the LOCOS citation is a placeholder pending the arXiv release.