Dataset card: document VITW columns
Browse files
README.md
CHANGED
|
@@ -19,6 +19,18 @@ dataset_info:
|
|
| 19 |
dtype: string
|
| 20 |
- name: description
|
| 21 |
dtype: string
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
splits:
|
| 23 |
- name: train
|
| 24 |
num_bytes: 150491
|
|
@@ -51,6 +63,16 @@ One row per value:
|
|
| 51 |
| `value` | short name, a slug of the Values in the Wild L1 cluster name |
|
| 52 |
| `cluster` | one of `Attunement`, `Rigor`, `Integrity`, `Stewardship` |
|
| 53 |
| `description` | the one-sentence value description the persona vectors were extracted from |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 54 |
|
| 55 |
```python
|
| 56 |
from datasets import load_dataset
|
|
@@ -75,8 +97,8 @@ Each cluster is named after its medoid value.
|
|
| 75 |
present-tense statement of how an assistant behaves when it holds the value (e.g.
|
| 76 |
`accessibility`: "using inclusive communication and representation to enable equal
|
| 77 |
participation and acceptance…"). The keys are slugs of the original L1 names. One
|
| 78 |
-
name appears under two L2 parents, so
|
| 79 |
-
added as a suffix.
|
| 80 |
2. **Persona vectors.** One persona vector per value, extracted from
|
| 81 |
OLMo-3.1-32B-Instruct-SFT with the
|
| 82 |
[persona vectors](https://github.com/safety-research/persona_vectors) recipe.
|
|
|
|
| 19 |
dtype: string
|
| 20 |
- name: description
|
| 21 |
dtype: string
|
| 22 |
+
- name: vitw_cluster_id
|
| 23 |
+
dtype: string
|
| 24 |
+
- name: vitw_name
|
| 25 |
+
dtype: string
|
| 26 |
+
- name: vitw_description
|
| 27 |
+
dtype: string
|
| 28 |
+
- name: vitw_l2
|
| 29 |
+
dtype: string
|
| 30 |
+
- name: vitw_l3
|
| 31 |
+
dtype: string
|
| 32 |
+
- name: vitw_pct_occurrences
|
| 33 |
+
dtype: float64
|
| 34 |
splits:
|
| 35 |
- name: train
|
| 36 |
num_bytes: 150491
|
|
|
|
| 63 |
| `value` | short name, a slug of the Values in the Wild L1 cluster name |
|
| 64 |
| `cluster` | one of `Attunement`, `Rigor`, `Integrity`, `Stewardship` |
|
| 65 |
| `description` | the one-sentence value description the persona vectors were extracted from |
|
| 66 |
+
| `vitw_cluster_id` | the value's Values in the Wild L1 `cluster_id`, the exact join key into upstream `values_tree.csv` |
|
| 67 |
+
| `vitw_name` | the original L1 cluster name |
|
| 68 |
+
| `vitw_description` | Anthropic's original L1 description, which `description` was rewritten from |
|
| 69 |
+
| `vitw_l2` | the L1 cluster's L2 parent |
|
| 70 |
+
| `vitw_l3` | the L3 top-level category (Practical, Epistemic, Social, Protective or Personal values) |
|
| 71 |
+
| `vitw_pct_occurrences` | upstream `pct_total_occurrences` for the L1 cluster |
|
| 72 |
+
|
| 73 |
+
Join on `vitw_cluster_id`, not on names: two different L1 clusters are both named
|
| 74 |
+
"Environmental stewardship". To get the leaf values inside a cluster, select the rows of
|
| 75 |
+
upstream `values_tree.csv` whose `parent_cluster_id` equals `vitw_cluster_id`.
|
| 76 |
|
| 77 |
```python
|
| 78 |
from datasets import load_dataset
|
|
|
|
| 97 |
present-tense statement of how an assistant behaves when it holds the value (e.g.
|
| 98 |
`accessibility`: "using inclusive communication and representation to enable equal
|
| 99 |
participation and acceptance…"). The keys are slugs of the original L1 names. One
|
| 100 |
+
name ("Environmental stewardship") appears under two L2 parents, so both copies
|
| 101 |
+
have their L2 parent's slug added as a suffix.
|
| 102 |
2. **Persona vectors.** One persona vector per value, extracted from
|
| 103 |
OLMo-3.1-32B-Instruct-SFT with the
|
| 104 |
[persona vectors](https://github.com/safety-research/persona_vectors) recipe.
|