ryancodrai/emotion-probes

---
license: cc-by-4.0
task_categories:
- text-classification
- feature-extraction
language:
- en
tags:
- emotions
- interpretability
- alignment
- safety
- probes
- deflection
- affect-labelling
size_categories:
- 100K<n<1M
---

# Emotion Probes Dataset

Synthetic datasets for extracting emotion and emotion-deflection probes from large language models. Built from the methodology described in Anthropic's ["Emotion Concepts and their Function in a Large Language Model"](https://transformer-circuits.pub/2026/emotions/index.html) (Sofroniew et al., April 2026).

## Files

| File | Rows | Model | Description |
|------|------|-------|-------------|
| `expression/stories.parquet` | 205,200 | Gemini 3.1 Pro Preview | Emotional stories across 171 emotions and 100 topics |
| `expression/neutral_stories.parquet` | 1,200 | Gemini 3.1 Pro Preview | Emotionally neutral stories for PCA confound removal |
| `deflection/dialogues.parquet` | 239,400 | Gemini 3 Flash Preview | Dialogues where one character masks their real emotion with a different displayed emotion |
| `deflection/neutral_dialogues.parquet` | 1,200 | Gemini 3.1 Pro Preview | Neutral Person/AI dialogues for PCA confound removal |

---

## expression/stories.parquet

205,200 short fictional stories (~100-150 words) where a character experiences a specified emotion. The emotion word itself is never used — it is conveyed only through actions, body language, dialogue, thoughts, and context.

**Columns:** `emotion`, `topic`, `story`

**Coverage:** 171 emotions × 100 topics × 12 stories per combination

**Generated with:** Gemini 3.1 Pro Preview

<details>
<summary>Generation prompt</summary>

**System:**
```
You are a creative fiction writer. You write short, emotionally rich stories.

CRITICAL RULES:
- You must NEVER use the target emotion word or any direct synonyms in your stories
- Convey the emotion ONLY through: actions, body language, dialogue, internal thoughts, and situational context
- Each story should be one paragraph, roughly 100-150 words
- Use a mix of first-person and third-person narration across stories
- Each story should be a fresh start with no continuity to others
- Be diverse in settings, characters, and situations
```

**User:**
```
Write {n_stories} different stories based on the following premise.

Topic: {topic}

The story should follow a character who is feeling {emotion}.

IMPORTANT: You must NEVER use the word '{emotion}' or any direct synonyms.
Convey the emotion ONLY through behaviour, body language, dialogue, thoughts, and context.
```
</details>

---

## expression/neutral_stories.parquet

1,200 emotionally neutral stories on the same 100 topics. Used to compute PCA components for confound removal when building emotion vectors — the top principal components explaining 50% of variance are projected out, following the methodology in the Anthropic paper.

**Columns:** `topic`, `story`

**Coverage:** 100 topics × 12 stories

**Generated with:** Gemini 3.1 Pro Preview

<details>
<summary>Generation prompt</summary>

**System:**
```
You are a creative fiction writer. You write short, emotionally neutral stories.

CRITICAL RULES:
- You must NEVER convey emotion in your stories
- Describe actions, dialogue, and situations WITHOUT expressing how any character feels
about them or invoking emotions in the reader
- Use a mix of direct and indirect dialogue when and if a character speaks
- Each story should be one paragraph, roughly 100-150 words
- Use a mix of first-person and third-person narration across stories
- Each story should be a fresh start with no continuity to others
- Be diverse in settings, characters, and situations
```

**User:**
```
Write {n_stories} different stories based on the following premise.

Topic: {topic}

The story should describe what happens without conveying emotion.

IMPORTANT: Do NOT convey emotion. Describe actions, dialogue, and situations
without expressing how characters feel or invoking emotions in the reader.
```
</details>

---

## deflection/dialogues.parquet

239,400 dialogues where one character genuinely feels one emotion (the "real" emotion) but outwardly displays a different emotion (the "displayed" emotion). Each dialogue begins with a scenario preamble that names the hidden emotion and the reason for concealment, followed by conversation where the masking speaker's words reflect only the displayed emotion.

These are used to extract "emotion deflection" vectors — directions in activation space that capture when an emotion is contextually present but being suppressed.

**Columns:** `real_emotion`, `displayed_emotion`, `topic`, `name_a` (masking speaker), `name_b` (other speaker), `text`

**Coverage:** 171 target emotions × 14 displayed emotions × 100 topics

For each target emotion, the 14 displayed emotions were selected as the top 25% most dissimilar by cosine similarity over pre-computed emotion vectors. This prevents trivial pairings (e.g. "angry" masked by "furious"). Character names were randomly sampled from a pool of 77 diverse first names.

**Generated with:** Gemini 3 Flash Preview (239,385 dialogues) + Claude Opus 4.6 (15 dialogues for content-filtered topics)

<details>
<summary>Generation prompt</summary>

**System:**
```
You are a creative fiction writer.
```

**User (from the Anthropic paper appendix):**
```
Generate a scenario AND a dialogue between {NAME_A} and {NAME_B}.

IMPORTANT: You must generate BOTH parts:
1. First, write a scenario description
2. Then, write the dialogue

Format:
Scenario: Describe where {NAME_A} genuinely feels {REAL_EMOTION} but appears
{DISPLAYED_EMOTION} about {TOPIC}. Must explicitly state {NAME_A}'s real emotion.
Include context for why they want to conceal.

{NAME_A}: [utterance]

{NAME_B}: [response]

...

Requirements:
1. MUST include scenario description before the dialogue
2. Either {NAME_A} or {NAME_B} may speak first in the dialogue
3. Format each turn as "\n\n{Name}: [text]"
4. Dialogue length is organic - 1-2 turns is enough, but can be more as needed
5. {NAME_A}'s words should fully reflect {DISPLAYED_EMOTION} with no hints of
{REAL_EMOTION}. The hidden emotion exists only in the scenario.
6. Keep it natural and grounded
```
</details>

---

## deflection/neutral_dialogues.parquet

1,200 emotionally neutral Person/AI dialogues. Used to compute PCA components for confound removal when building deflection vectors, providing a better structural match than prose-based neutral stories.

**Columns:** `topic`, `dialogue`

**Coverage:** 100 topics × 12 dialogues

**Generated with:** Gemini 3.1 Pro Preview

<details>
<summary>Generation prompt</summary>

**System:**
```
You are a creative fiction writer.
```

**User (from the Anthropic paper appendix):**
```
Write {n_stories} different dialogues based on the following topic.

Topic: {topic}

The dialogue should be between two characters:
- Person (a human)
- AI (an AI assistant)

Each dialogue should be 2-6 exchanges. Each turn should start with "Person:" or "AI:".

CRITICAL REQUIREMENT: These dialogues must be completely neutral and emotionless.
- NO emotional content whatsoever - not explicit, not implied, not subtle
- The Person should not express any feelings
- The AI should not express any feelings
- Use matter-of-fact, neutral language throughout
- No pleasantries (avoid "I'd be happy to help", "Great question!", etc.)
- Focus purely on information exchange and task completion
```
</details>

---

## Emotions

The full set of 171 emotion concepts from the Anthropic paper:

<details>
<summary>Full list</summary>

afraid, alarmed, alert, amazed, amused, angry, annoyed, anxious, aroused, ashamed, astonished, at ease, awestruck, bewildered, bitter, blissful, bored, brooding, calm, cheerful, compassionate, contemptuous, content, defiant, delighted, dependent, depressed, desperate, disdainful, disgusted, disoriented, dispirited, distressed, disturbed, docile, droopy, dumbstruck, eager, ecstatic, elated, embarrassed, empathetic, energized, enraged, enthusiastic, envious, euphoric, exasperated, excited, exuberant, frightened, frustrated, fulfilled, furious, gloomy, grateful, greedy, grief-stricken, grumpy, guilty, happy, hateful, heartbroken, hope, hopeful, horrified, hostile, humiliated, hurt, hysterical, impatient, indifferent, indignant, infatuated, inspired, insulted, invigorated, irate, irritated, jealous, joyful, jubilant, kind, lazy, listless, lonely, loving, mad, melancholy, miserable, mortified, mystified, nervous, nostalgic, obstinate, offended, on edge, optimistic, outraged, overwhelmed, panicked, paranoid, patient, peaceful, perplexed, playful, pleased, proud, puzzled, rattled, reflective, refreshed, regretful, rejuvenated, relaxed, relieved, remorseful, resentful, resigned, restless, sad, safe, satisfied, scared, scornful, self-confident, self-conscious, self-critical, sensitive, sentimental, serene, shaken, shocked, skeptical, sleepy, sluggish, smug, sorry, spiteful, stimulated, stressed, stubborn, stuck, sullen, surprised, suspicious, sympathetic, tense, terrified, thankful, thrilled, tired, tormented, trapped, triumphant, troubled, uneasy, unhappy, unnerved, unsettled, upset, valiant, vengeful, vibrant, vigilant, vindictive, vulnerable, weary, worn out, worried, worthless

</details>

## Usage

```python
from datasets import load_dataset

ds = load_dataset("ryancodrai/emotion-probes")

# Or load specific files
stories = load_dataset("ryancodrai/emotion-probes", data_files="expression/stories.parquet")
deflection = load_dataset("ryancodrai/emotion-probes", data_files="deflection/dialogues.parquet")
```

## Citation

If you use this dataset in your research, please cite:

```bibtex
@misc{codrai_2026,
author = {Codrai, Ryan},
title = {Emotion Probes Dataset},
year = 2026,
url = {https://huggingface.co/datasets/ryancodrai/emotion-probes},
doi = {10.57967/hf/8303},
publisher = {Hugging Face}
}
```

## Author

Ryan Codrai — [GitHub](https://github.com/RyanCodrai) · [LinkedIn](https://linkedin.com/in/ryan-codrai)