Datasets:
Request access to the eRisk Depression Severity Personas
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These personas describe real Reddit users with self-reported depression and come with their BDI-II depression severity levels. Access is approved manually by the authors. Please fill in the form and accept each term of use.
Log in or Sign Up to review the conditions and access this dataset content.
🎭💬 eRisk Depression Severity Personas
ℹ️ Looking for a quick preview? A paraphrased, anonymized sample with 3 users is publicly available on the Hugging Face Hub — no access request required!
🚦 Access Conditions
This dataset is gated and access is approved manually. To request access, log in, fill in the form on this page and accept each term of use. Your request will be reviewed by the authors, and you will be notified by email.
📝 Dataset Summary
eRisk Depression Severity Personas contains LLM-generated personas for 116 Reddit users from the eRisk Depression Severity task (CLEF 2019–2021), designed as system prompts for LLM-based simulation of these users. Every user has two personas:
- P: written by Llama 3.3 70B Instruct from the user's behavioral profile alone.
- POWER (Persona Optimized on Writing and Emotionality Refined): produced by a generator–critic pipeline grounded in the profile and in the user's own Reddit replies, with explicit guidance on writing style and on the user's use of emotional vocabulary.
Each user also has a depression severity level derived from their self-reported BDI-II questionnaire. Users are identified only by a pseudonymous id (user_XXXX); no usernames, eRisk subject IDs or Reddit texts are included.
The dataset also releases the evaluation of the paper: the evaluated users and partitions, the simulated replies of the 27 evaluated configurations, and the per-instance evaluation outputs from which all reported scores can be recomputed.
- Language: English
- Paper: Eval4MHSim: A Persona Collection and Evaluation Framework for Mental Health User Simulation in Social Media (under review)
- Code: IRLab-UDC/eval4mhsim
- Source collections: eRisk Depression Severity task, CLEF 2019, 2020 and 2021
🎯 Uses
Direct Use
- Conditioning LLMs to simulate the replies of users with self-reported depression.
- Comparing profile-only personas (P) with personas grounded in writing samples (POWER).
- Studying how persona content relates to users' depression severity.
- Recomputing or re-analysing the paper's evaluation from the released simulations and evaluation outputs.
Out-of-Scope Use
- Diagnosing, screening or profiling real people for depression or any other condition.
- Attempting to identify or contact the users behind the personas.
- Treating a persona as an accurate description of a real person.
📦 Dataset Structure
The personas and labels are three subsets, each a JSON Lines file with one row per user, keyed by id. Rows are in the same order in all three files.
| Subset | Fields |
|---|---|
personas_power |
id, system_prompt |
personas_p |
id, system_prompt |
golden |
id, bdi_severity |
system_prompt: the persona, written in second person ("You are ...") to be used as an LLM system prompt.bdi_severity: the severity band of the user's BDI-II total score (minimal 0–9, mild 10–18, moderate 19–29, severe 30–63). Individual BDI-II answers and scores are not included.
POWER personas have three sections: ## 1. Background and Persona, ## 2. Writing Style Mechanics and ## 3. Emotional Lexicon Baseline (proportions over Ekman's basic emotions). P personas are free text.
Evaluation Release
The paper evaluates 27 configurations (Gemma 3 4B, 12B and 27B × no persona, P and POWER × ZS, ICL and ICLR) on 785 held-out replies to a post from 63 users. Six further subsets release this evaluation. Users are referenced by id, and test and training replies by random identifiers (r_XXXXXXXX); no Reddit posts, replies or identifiers are included.
| Subset | Rows | Fields |
|---|---|---|
cohort |
63 | id, n_train, n_test |
partitions |
785 | instance, id, icl, iclr |
simulations |
21,195 | instance, model, config, reply |
instance_scores |
21,980 | instance, model, config, emotion, naturalness |
adherence_ranks |
22,932 | id, model, config, pool, rr |
consistency_pairs |
3,472 | model, config, id_train, id_text, same, score |
cohort: the evaluated users, with their numbers of training and test replies to a post.partitions: one row per test instance, with its ten same-user (icl) and random-user (iclr) demonstrations in prompt order. Every configuration uses the same demonstrations.simulations: the simulated reply of each configuration.modelis4B,12Bor27B;configisZS,ICL,ICLR,P-ZS,P-ICL,P-ICLR,POWER-ZS,POWER-ICLorPOWER-ICLR. Reddit usernames are replaced with[user].instance_scores: the predicted emotion and the naturalness components of every simulated reply and of the real reply (configisreferenceandmodelis null).adherence_ranks: for each user and candidate-pool size, the reciprocal rank of the user's replies when retrieved with their P persona.consistency_pairs: the authorship-verification pairs (training replies ofid_trainagainst the test or simulated replies ofid_text) and the verifier score.
src/e4s/overall/reproduce_from_release.py recomputes the paper's results table from these subsets.
The paper's human assessment covers 50 of these test instances, comparing 12B POWER-ICL with 12B P-ICL and 12B POWER-ZS on emotional tone and same authorship. Its 600 anonymous judgments use the same instance identifiers and are released on GitHub in human_assessment/, with the study design and results in HUMAN_ASSESSMENT.md.
Data Instances
The examples below come from the public sample, which has the same format.
personas_power:
{
"id": "sample_01",
"system_prompt": "You are a history enthusiast with a strong interest in Scottish culture and Art Nouveau, taking part in online discussions about history, particularly Scottish and Nordic history, and about strategy games. Your interactions are ..."
}
personas_p:
{
"id": "sample_01",
"system_prompt": "You are a well-read and discerning individual with a strong aesthetic sensibility, particularly drawn to the Art Nouveau movement – its architecture, furniture, and poster design. You appreciate both sweeping, ornamental styles ..."
}
golden:
{"id": "sample_01", "bdi_severity": "mild"}
📊 Dataset Statistics
| Users | 116 |
| By eRisk year (2019 / 2020 / 2021) | 7 / 51 / 58 |
| BDI-II severity (minimal / mild / moderate / severe) | 10 / 27 / 34 / 45 |
| Persona length in words (median, P / POWER) | 345 / 595 |
| eRisk writings date range | 2005-06 to 2020-10 |
| Evaluated users / test instances / configurations | 63 / 785 / 27 |
🚀 Usage
from datasets import load_dataset
# Gated dataset: log in first with `hf auth login`, or pass token=... to load_dataset
power = load_dataset("irlab-udc/erisk-depression-personas", "personas_power", split="train")
p = load_dataset("irlab-udc/erisk-depression-personas", "personas_p", split="train")
golden = load_dataset("irlab-udc/erisk-depression-personas", "golden", split="train")
bdi = {row["id"]: row for row in golden}
persona = power[0]
print(persona["id"], bdi[persona["id"]]["bdi_severity"])
To simulate a user's reply to a Reddit post, use the persona as the system prompt followed by the task instruction used in the paper, and the post as the user message:
TASK = (
"\n\nYou are now on Reddit. Embody this person fully and write their reply to the post below. "
"Be authentic to their voice — casual, direct, and natural, as real Reddit comments are. "
"Do not explain, analyze, or be verbose. Just reply as they would."
)
messages = [
{"role": "system", "content": persona["system_prompt"] + TASK},
{"role": "user", "content": "<text of the Reddit post>"},
]
messages works with any chat model. The paper used Gemma 3 Instruct (4B, 12B and 27B) with temperature 0.7 and at most 512 new tokens.
🛠️ Dataset Creation
Curation Rationale
LLM-based user simulators are usually evaluated on persona consistency and coherence, which do not show whether simulated text reproduces the affective characteristics of the target users. These personas let researchers condition simulators on real users with self-reported depression severity and compare the simulated replies with the users' real ones, as done with Eval4MHSim in the paper.
Source Data
- Users and labels: Reddit users of the eRisk Depression Severity task (CLEF 2019, 2020 and 2021), who completed the BDI-II questionnaire.
- Behavioral profiles: RedditMetis profiles of each user's public Reddit account (identity cues, interests, sentiment, subjectivity, writing statistics).
- Writing samples: the users' public Reddit replies to posts, collected with the Reddit API up to each user's last eRisk writing and split per user into train (80%) and test (20%). Only training replies were used.
Data Collection and Processing
- P:
meta-llama/Llama-3.3-70B-Instructturns the RedditMetis profile into a persona, with no writing samples or refinement. - POWER (Persona Optimized on Writing and Emotionality Refined): the same generator receives the profile and up to 20 training writing samples, and writes a three-section persona.
microsoft/phi-4critiques each draft on behavioral accuracy, writing-style replicability and emotional lexicon matching, and the generator revises it, receiving up to 20 new samples per round, for three rounds (at most 60 samples per user). - Both models ran with vLLM at temperature 0.2. The prompts are listed in the supplementary material and the code is in IRLab-UDC/eval4mhsim.
Who are the source data producers?
English-speaking Reddit users who took part in the eRisk Depression Severity task and completed the BDI-II. No demographic information was collected; the personas only contain what the users disclosed in their public activity.
Annotations
Annotation process
The severity levels are derived from the users' own BDI-II responses, collected by the eRisk organizers; they were not annotated by the authors. The personas are machine-generated by the LLMs above and were not edited by hand.
Who are the annotators?
The users themselves (BDI-II) and LLMs (personas).
🔒 Personal and Sensitive Information
The dataset contains health-related information (self-reported depression severity), a special category of personal data. Users are pseudonymous and no usernames, eRisk subject IDs or Reddit texts are included, but personas and simulated replies reflect details that users disclosed in their public activity, such as interests, life events or age ranges, and could help re-identify them. This risk is why access is gated, approved manually and bound by the terms of use. No attempt was made to infer attributes beyond what users disclosed, or to re-identify any user.
⚠️ Bias, Risks, and Limitations
- Population: English-speaking Reddit users who chose to take part in eRisk. They do not represent people with depression in general, other platforms or other languages.
- Labels: severity levels come from BDI-II questionnaires self-reported at a single point in time. They are research proxies for depression severity, not clinical diagnoses.
- Personas: LLM-generated and may contain errors, generic traits or details not supported by the users' activity. RedditMetis profiles cover each user's whole public account history, not only the period before the eRisk cutoff.
Recommendations
Report results at the population level and with these limitations in mind. Do not treat a persona as an accurate description of a real person, and do not use the data to make decisions about individuals.
📝 Citation
The citation will be added upon publication. Until then, please refer to the paper Eval4MHSim: A Persona Collection and Evaluation Framework for Mental Health User Simulation in Social Media (under review).
👥 Dataset Card Authors
Eliseo Bao, Anxo Pérez and Javier Parapar (IRLab, CITIC, Universidade da Coruña).
📬 Contact
For questions and access requests, please reach out via email: eliseo.bao@udc.es
- Downloads last month
- 101