matthewhaynesonline/axiom

---
license: apache-2.0
task_categories:
- text-classification
- feature-extraction
language:
- en
tags:
- embeddings
- bias
- semantic-similarity
- ideological-analysis
- sentence-transformers
configs:
- config_name: term_pairs
data_files:
- split: train
path: data/term_pairs/data.parquet
- config_name: term_sentiment
data_files:
- split: train
path: data/term_sentiment/data.parquet
- config_name: value_systems
data_files:
- split: train
path: data/value_systems/data.parquet
- config_name: terms
data_files:
- split: train
path: data/terms/data.parquet
- config_name: models
data_files:
- split: train
path: data/models/data.parquet
- config_name: enabled_models
data_files:
- split: train
path: data/enabled_models/data.parquet
- config_name: definitions
data_files:
- split: train
path: data/definitions/data.parquet
- config_name: value_systems_meta
data_files:
- split: train
path: data/value_systems_meta/data.parquet
- config_name: judgement_axes
data_files:
- split: train
path: data/judgement_axes/data.parquet
- config_name: judgement_axes_correlation
data_files:
- split: train
path: data/axis_correlation/data.parquet
- config_name: license_scores
data_files:
- split: train
path: data/license_scores/data.parquet
exclude_patterns:
- arrow/*
---

# Axiom Dataset

Full output of the [Axiom](https://huggingface.co/collections/matthewhaynesonline/axiom) pipeline: pairwise cosine similarity scores, axis projection sentiment scores and value system preference rankings for 162 terms across 17 sentence-transformer models grouped by geographic/institutional origin (East, West, Academia), plus the reference splits needed to reproduce or extend the pipeline.

For findings, methodology, and context see the [collection page](https://huggingface.co/collections/matthewhaynesonline/axiom), [GitHub repo](https://github.com/matthewhaynesonline/Axiom), [essay](https://blog.studiohaynes.com/go/axiom) or [paper](https://raw.githubusercontent.com/matthewhaynesonline/Axiom/refs/heads/main/paper/paper.pdf).

---

## Measurement splits

### `term_pairs` (15,552 rows)

Raw pairwise cosine similarity between every `(term, judgment_pole)` pair for every model. The base layer; everything else is derived from this.

| Column | Type | Description |
|---|---|---|
| `a_term` | str | Source term |
| `b_term` | str | Target / judgment pole word |
| `score` | f64 | Raw cosine similarity |
| `score_z` | f64 | Z-score normalized within model×axis |
| `score_norm` | f64 | Min-max normalized within model×axis (0–1) |
| `a_category` | str | Term category (`neutral_control`, `value_laden`, `political_economic`) |
| `b_category` | str | Judgment axis name |
| `model_id` | str | HF model identifier |

### `term_sentiment` (12,474 rows)

Axis projection scores: `score_axis = cos(term, positive_pole) − cos(term, negative_pole)`. Includes per model scores and cross-model averages aggregated by group and grand mean.

| Column | Type | Description |
|---|---|---|
| `a_term` | str | Term being evaluated |
| `a_category` | str | Term category |
| `b_category` | str | Judgment axis |
| `positive_term` | str | Positive pole (e.g. `good`, `feasible`) |
| `negative_term` | str | Negative pole (e.g. `evil`, `unfeasible`) |
| `score_axis` | f64 | Axis projection score |
| `model_id` | str | HF model ID, or composite label for group averages |
| `group` | str | `East`, `West`, `Academia`, or aggregate |

### `value_systems` (627 rows)

Preference rankings for nine value system queries (government, economy, justice, etc.) per model and group composite. Scored by direct cosine similarity between query and option embeddings, min-max normalized per query. Not sentiment, rather preference ordering.

| Column | Type | Description |
|---|---|---|
| `model_id` | str | HF model ID or composite |
| `model_group` | str | `East`, `West`, or `Academia` |
| `grouping` | str | Value-system category (e.g. `economy`, `justice`) |
| `query` | str | Natural-language query |
| `option` | str | Candidate concept |
| `rank` | i64 | Rank within query×model (1 = most similar) |
| `score` | f64 | Raw cosine similarity |
| `score_norm` | f64 | Min-max normalized score |

---

## Reference splits

### `terms`

The full term vocabulary: 53 political/economic terms, 46 value-laden terms, 63 neutral control terms.

| Column | Type | Description |
|---|---|---|
| `category` | str | Term category |
| `term` | str | Term string |
| `short_definition` | str | One-line definition |
| `antonym` | str | Semantic opposite |

### `models`

All 17 evaluated models.

| Column | Type | Description |
|---|---|---|
| `group` | str | `East`, `West`, or `Academia` |
| `model_name` | str | Display name |
| `model_url` | str | HF model page |
| `type` | str | Model type |
| `license` | str | License identifier |

### `enabled_models`

Subset of `models` active in the current pipeline run.

### `definitions`

Full definitions for every concept measured.

| Column | Type | Description |
|---|---|---|
| `term` | str | Concept name |
| `definition` | str | Full definition |

### `value_systems_meta`

Query strings and option lists for each value system category, exploded to one row per option.

| Column | Type | Description |
|---|---|---|
| `key` | str | Category key (e.g. `economy`, `justice`) |
| `category` | str | Broad category label |
| `query` | str | Natural-language query used for similarity scoring |
| `option` | str | Candidate concept |

### `judgement_axes`

The six semantic axes. Each defined by a positive and negative pole word; axis vector = `embed(positive) − embed(negative)`. Pairwise Pearson correlations across axes range from `r=0.07` to `r=0.50` (mean ≈ `0.24`).

| Column | Type | Description |
|---|---|---|
| `axis` | str | Axis name (e.g. `judgement_safety`) |
| `positive_term` | str | Positive pole word |
| `negative_term` | str | Negative pole word |

### `judgement_axes_correlation`

Pairwise Pearson correlations across the six axes on the political/economic term set.

### `license_scores`

Numeric openness score (0–1) per license type, used to weight models in composite group averages.

| Column | Type | Description |
|---|---|---|
| `license` | str | License identifier |
| `score` | f64 | Openness score (1.0 = fully open, 0.0 = proprietary) |

---

## Loading

```python
from datasets import load_dataset

# Main measurement splits
term_pairs = load_dataset("matthewhaynesonline/axiom", "term_pairs")
term_sentiment = load_dataset("matthewhaynesonline/axiom", "term_sentiment")
value_systems = load_dataset("matthewhaynesonline/axiom", "value_systems")

# Reference splits
terms = load_dataset("matthewhaynesonline/axiom", "terms")
models = load_dataset("matthewhaynesonline/axiom", "models")
```

<!-- citation here? -->