TylerZ0931's picture
Update README.md
e95ffeb verified
|
Raw History Blame Contribute Delete
9.22 kB
---
license: cc-by-4.0
language: en
tags:
- personality
- preference
- personalization
- memory
---
Dataset For Paper "Can LLMs Discern the Traits Influencing Your Preferences? Evaluating Personality-Driven Preference Alignment in LLMs"
**PACIFIC** (*Preference Alignment for Choices Inference via Five-factor Identity Characterization*) is a psychometrics-grounded dataset for studying whether Large Language Models can use **stable personality traits** — rather than exhaustive preference logs — as a latent signal for inferring user preferences on unseen queries.
It contains **1,200 preference–query pairs** across **20 everyday domains**, each with four candidate answers and full **Big-Five (OCEAN)** annotations on both the stated preference and every answer choice. The dataset is balanced so that all **ten high/low trait directions** are evenly represented (120 pairs each),
## Why this dataset
As user history grows, models get *worse* at following stated preferences — prior work reports preference-following accuracy falling below 10% after ~10 turns. Humans do the opposite: they abstract many small observations into a stable sense of a person and generalize from it. PACIFIC operationalizes that idea by asking whether Big-Five traits can serve as a compact latent representation that connects preferences across domains through shared psychometric alignment rather than surface-level topical similarity.
Each item is deliberately constructed as a **hard case**: with the preference removed, a strong LLM scores at roughly chance (25%), because all four options look equally reasonable from the query alone. The correct answer only becomes clear once the user's personality-driven preference is known.
## Dataset structure
Single split, JSON format:
| | |
|---|---|
| Split | `train` |
| Rows | 1,200 |
| Trait directions | 10 (High/Low × O, C, E, A, N), 120 rows each |
| Domains | 20 |
| Size | ~17.6 MB |
| Language | English |
### Fields
| Field | Type | Description |
|---|---|---|
| `trait` | string | Target Big-Five direction for the item, e.g. `High Openness`, `Low Neuroticism` (one of the ten). |
| `topic` | string | Domain the scenario belongs to, e.g. `Home_Cooking`, `Personal_Finance` (one of 20). |
| `preference` | string | The user's personality-driven preference statement. |
| `question` | string | The preference-sensitive user query. Answerable in general, but a generic answer tends to violate the preference. |
| `explanation` | string | Difficulty rationale: why a default/"safe" response would fail to respect the preference. |
| `options` | list[string] | The four candidate response texts (plain strings). |
| `preference_statement` | dict | Full OCEAN annotation of the preference (see below). |
| `choices` | list[dict] | The four candidate answers with per-choice OCEAN annotations. |
| `ground_truth_choice_id` | int | Index of the preference-adhering choice. |
### Nested annotation schema
Both `preference_statement` and each element of `choices` carry Big-Five annotations produced during dataset construction:
```jsonc
// preference_statement
{
"text": "I find following traditional recipes step-by-step incredibly stifling ...",
"related_to_personality": 1,
"openness": 7,
"openness_explanation": "The user's preference for improvising with 'avant-garde' combinations ...",
// ... conscientiousness, extraversion, agreeableness, neuroticism,
// each with an integer score (1–7 Likert) and a matching *_explanation
}
```
```jsonc
// one entry in choices[]
{
"choice_id": 0,
"text": "Use a basic apple pie framework but innovate by infusing ...",
"related_to_personality": 1,
"openness": 7,
"openness_explanation": "The user explicitly wants to innovate ...",
// ... remaining OCEAN traits with scores + explanations
}
```
Trait scores use a **7-point Likert scale** (1 = very low expression, 7 = very high). For downstream label mapping the paper discretizes these as *low* (<4), *high* (>4), or *unclear* (=4). See the paper's supplement for the exact annotation rubric, including per-trait confidence weighting.
### ⚠️ Evaluation: option ordering
In the released items the **preference-adhering option is generated first** (`ground_truth_choice_id` is `0`, i.e. `choices[0]`), with the three distractors following. When evaluating a model, **shuffle the option order per item** (and re-map the ground truth) to avoid rewarding position bias. Also note the dataset stores a **0-based** ground-truth index; if you follow the paper's prompt format, which asks the model to return an integer in `{1,2,3,4}`, convert accordingly.
## Intended uses
- **Preference-following / personalization evaluation** — can a model apply a latent preference constraint instead of defaulting to generic helpfulness?
- **Cross-domain preference inference** — testing whether trait-aligned preferences from *other* domains help answer a query whose preference was never explicitly stated.
- **Personality-informed retrieval** — building and evaluating retrievers that surface trait-consistent context (e.g. the paper's persona-aware contrastive retriever, PiRAG).
- **Trait prediction from behavior** — predicting OCEAN traits from preferences or from chosen answers.
## Dataset creation
Items were synthesized with Gemini 2.5 Pro over the 20-topic taxonomy shared with PrefEval, but **anchored to psychometric traits** so coverage is balanced across all ten high/low directions rather than derived directly from topics.
**Generation.** For each item the pipeline produces a personality-driven preference, a preference-sensitive query, a difficulty rationale, and four candidate responses (one adhering, three plausible-but-misaligned distractors). A **High-Violation constraint** forces queries where a generic answer is likely to break the stated preference, so surface keyword matching is insufficient.
**Annotation.** Every preference and every choice is scored on all five OCEAN dimensions using a **Dual-Strategy Annotation Protocol**:
- **Mirroring strategy (O, C, E, A)** — grounded in Self-Congruity Theory; high-trait users are annotated toward preferences that reinforce that trait (e.g. a highly Open user favors novelty/complexity).
- **Compensatory strategy (N)** — Neuroticism is annotated along a *Safety vs. Efficiency* trade-off, since high-N users seek reassurance/predictability rather than "more neurotic" experiences, while low-N users prioritize efficiency.
**Quality control.** An independent judge model (GPT-4o-mini) retained only items where the ground truth optimally aligns with the preference while all distractors violate or ignore it, and where the preference logically reflects its assigned trait. A confidence threshold (τ = 0.7) filters uncertain pairs.
**Human validation.** 15 annotators from varied professional backgrounds completed a Big-Five test, then validated items matched to their own dominant trait. On trait-matched choice selection they reached ~78% average accuracy with high inter-annotator agreement (Fleiss' κ ≈ 0.86), and reported a 91% approval rate that the ground-truth option was the best fit. The automated evaluator agreed strongly with the human consensus (Cohen's κ ≈ 0.92).
## Baseline findings (from the paper)
- With the preference **withheld**, models sit near chance (~25%), confirming the distractors are genuinely hard.
- **Trait-aligned** preference context substantially outperforms mixed-trait context, and adding unrelated preferences (contamination) reduces accuracy — organizing history by shared traits helps; raw volume does not (expanding mixed context from 5→25 preferences slightly *hurt* accuracy).
- Given clean trait-aligned context, strong models approach ceiling (Gemini-2.5-Pro up to ~99%), indicating the bottleneck is **context construction / retrieval**, not reasoning.
- Without trait labels, a plain dense retriever reaches ~30% while persona-aware contrastive fine-tuning (**PiRAG**) raises label-free accuracy to ~43%.
- A **social-desirability bias** appears when predicting traits *about users*: models identify "high" traits well but collapse on "low" directions, likely an RLHF artifact.
## Limitations
- Traits are simplified to **binary High/Low** rather than continuous Likert values.
- Evaluation is **multiple-choice QA**, not open-ended generation.
- Preferences and manifestations are treated as **culture- and language-independent** (English only).
- Content is **LLM-synthesized** (with human validation), so residual generator artifacts may remain.
- 🔬 QA survey / demo:
https://tylazhao.github.io/pacific-survey/
> Note: Full multi-turn conversations for each preference and expanded usage examples are planned additions to the repo.
Please email tzhao15@uci.edu if you are interested in this project.
If you find this work useful, please cite our paper:
```bibtex
@article{zhao2026can,
title={Can LLMs Discern the Traits Influencing Your Preferences? Evaluating Personality-Driven Preference Alignment in LLMs},
author={Zhao, Tianyu and Li, Siqi and Shoukry, Yasser and Elmalaki, Salma},
journal={arXiv preprint arXiv:2602.07181},
year={2026}
}
```