MMHusnain/semeval26-psysemev-detection-2026-05-17-1834

SemEval 2026 Task 10 — Conspiracy Detection

RoBERTa-large with LoRA for binary conspiracy detection on Reddit submission statements (SemEval 2026 Task 10, detection subtask).

Model description

• —Task: Classify each document as conspiracy-related (Yes) or not (No). Samples labeled Can't tell in training are excluded from the binary setup.
• —Architecture: roberta-large + PEFT LoRA adapter (sequence classification, 2 labels).
• —Labels: No (0), Yes (1).
• —Max length: 512 tokens.

Intended use

Research and evaluation on the SemEval 2026 psycholinguistic markers shared task. Not intended for high-stakes or real-time moderation without further validation.

How to load

from transformers import AutoModelForSequenceClassification, AutoTokenizer
from peft import PeftModel

repo_id = "MMHusnain/semeval26-psysemev-detection-2026-05-17-1834"
base_model = "roberta-large"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(base_model, num_labels=2)
model = PeftModel.from_pretrained(model, repo_id)

For merged weights (if you merge LoRA locally), load with AutoModelForSequenceClassification.from_pretrained(repo_id) only.

Training

Trained with parameter-efficient fine-tuning (LoRA) on rehydrated task data. See the project repository for splits, hyperparameters, and train_and_infer_binary.py / infer_binary.py.

Evaluation

Binary metrics on Yes / No dev labels (e.g. dev_public.jsonl). Use the companion extraction models for span-level psycholinguistic markers.

Limitations

• —Trained on English Reddit submission statements; may not generalize to other domains or languages.
• —Task labels have moderate annotator agreement; treat outputs as research signals, not ground truth.
• —Oversampled conspiracy-related content in the corpus — prevalence on open web may differ.

Citation

SemEval-2026 Task 10: Psycholinguistic Markers of Conspiracy Theories in Social Media Conversations.

Model metadata