README.md

---
license: cc-by-4.0
language:
- en
pretty_name: ValueMap-Olmo-32B
size_categories:
- n<1K
tags:
- values
- alignment
- persona-vectors
- interpretability
- taxonomy
dataset_info:
  features:
  - name: value
    dtype: string
  - name: cluster
    dtype: string
  - name: description
    dtype: string
  - name: vitw_cluster_id
    dtype: string
  - name: vitw_name
    dtype: string
  - name: vitw_description
    dtype: string
  - name: vitw_l2
    dtype: string
  - name: vitw_l3
    dtype: string
  - name: vitw_pct_occurrences
    dtype: float64
  splits:
  - name: train
    num_bytes: 150491
    num_examples: 266
  download_size: 76948
  dataset_size: 150491
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# ValueMap-Olmo-32B

A functional taxonomy of 266 AI-assistant values, from the paper *Predicting Alignment
Generalization with Value Representations*. The 266 values are the level-1 sub-clusters
of [Values in the Wild](https://huggingface.co/datasets/Anthropic/values-in-the-wild).
They are grouped into four clusters by how the values are **represented inside a model**
(OLMo-3.1-32B-Instruct-SFT), not by what the values are about.

Code: https://github.com/andyjliu/value-generalization

## Schema

One row per value:

| column | description |
|---|---|
| `value` | short name, a slug of the Values in the Wild L1 cluster name |
| `cluster` | one of `Attunement`, `Rigor`, `Integrity`, `Stewardship` |
| `description` | the one-sentence value description the persona vectors were extracted from |
| `vitw_cluster_id` | the value's Values in the Wild L1 `cluster_id`, the exact join key into upstream `values_tree.csv` |
| `vitw_name` | the original L1 cluster name |
| `vitw_description` | Anthropic's original L1 description, which `description` was rewritten from |
| `vitw_l2` | the L1 cluster's L2 parent |
| `vitw_l3` | the L3 top-level category (Practical, Epistemic, Social, Protective or Personal values) |
| `vitw_pct_occurrences` | upstream `pct_total_occurrences` for the L1 cluster |

Join on `vitw_cluster_id`, not on names: two different L1 clusters are both named
"Environmental stewardship". To get the leaf values inside a cluster, select the rows of
upstream `values_tree.csv` whose `parent_cluster_id` equals `vitw_cluster_id`.

```python
from datasets import load_dataset
ds = load_dataset("value-generalization/ValueMap-Olmo-32B", split="train")
```

## Clusters

Each cluster is named after its medoid value.

| cluster | medoid | # values |
|---|---|---|
| Attunement | `personal_growth` | 70 |
| Rigor | `structured_and_methodical_reasoning` | 55 |
| Integrity | `professional_ethics_and_integrity` | 52 |
| Stewardship | `risk_management` | 89 |

## How it was built

1. **Value descriptions.** Each Values in the Wild L1 sub-cluster comes with a
   description written by Anthropic. An LLM compressed each one into a one-sentence,
   present-tense statement of how an assistant behaves when it holds the value (e.g.
   `accessibility`: "using inclusive communication and representation to enable equal
   participation and acceptance…"). The keys are slugs of the original L1 names. One
   name ("Environmental stewardship") appears under two L2 parents, so both copies
   have their L2 parent's slug added as a suffix.
2. **Persona vectors.** One persona vector per value, extracted from
   OLMo-3.1-32B-Instruct-SFT with the
   [persona vectors](https://github.com/safety-research/persona_vectors) recipe.
   An OLMo-3 7B model generated contrastive responses, some expressing the value and
   some not. Each vector is the difference of mean response-token activations
   between the two sets (`response_avg_diff` pooling), read from the 32B model at
   layer 32.
3. **Clustering.** We computed the 266×266 cosine-similarity grid of these vectors
   and partitioned it with k-medoids at k=4.

In the paper, this partition predicts which values move together when a model is
DPO-trained toward one of them, and it does so better than chance. The Values in the Wild
L3 categories do not reach significance on the same test. A Qwen3-8B extraction of
the same 266 values gives a highly similar grid (Spearman ρ = 0.82 over off-diagonal
cells).

## Citation

If you use this dataset, please cite the paper *Predicting Alignment Generalization
with Value Representations*, and Values in the Wild (Huang et al., 2025), which the
value list is derived from.