P2P-DeepSeek-R1-8B-MBTI-LoRA

P2P: DeepSeek-R1-8B Fine-Tuned for MBTI Personality Prediction

A QLoRA fine-tuned adapter for predicting Myers-Briggs personality types (MBTI) from social media posts, based on the P2P framework.

Author: Omar Gamal ElKady | ITI - AI Track, Intake 46

Paper

Ma, T., Feng, K., Rong, Y., & Zhao, K. (2025). From Post To Personality: Harnessing LLMs for MBTI Prediction in Social Media. In Proceedings of CIKM '25. ACM.

Model Details

• Developed by: Omar Gamal ElKady
• Affiliation: Information Technology Institute (ITI), AI Track, Intake 46
• Model type: LoRA adapter for causal language model
• Language: English
• Base model: DeepSeek-R1-Distill-Llama-8B (4-bit quantized)
• Fine-tuning method: QLoRA via Unsloth
• License: MIT

Training Configuration

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model (4-bit)
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit",
    device_map="auto",
    torch_dtype="float16",
)

# Attach fine-tuned adapter
model = PeftModel.from_pretrained(base_model, "OmarGamal488/P2P-DeepSeek-R1-8B-MBTI-LoRA")
tokenizer = AutoTokenizer.from_pretrained("OmarGamal488/P2P-DeepSeek-R1-8B-MBTI-LoRA")

# Predict MBTI
prompt = """Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
Predict the 4-letter Myers-Briggs personality type of the user from their social-media posts. Return only four uppercase letters.

### Input:
love debate theory logic prefer alone question everything analyze pattern fascinated philosophy

### Response:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=10, do_sample=False)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
# Output: INTP

Pipeline Overview

This model is the local LLM component of the P2P framework:

User Posts --> [This Model] --> Personality Features + Hidden States
                                           |
                                           v
                                    FAISS (k=5 RAG)
                                           |
                                           v
                               [DeepSeek-V3 API] --> 4-letter MBTI Type

The model serves two roles in the pipeline:

1. Feature Extraction — Reads user posts and generates a structured personality assessment covering E/I, S/N, T/F, J/P dimensions
2. Hidden State Encoding — Produces 4096-dim internal representations used for FAISS similarity search

Training Data

PersonalityCafe forum dataset from Kaggle:

• 8,675 users with self-reported MBTI labels
• 50 most recent posts per user
• Preprocessing: lowercase, URL removal, stopword removal, lemmatization
• Split: 60% train / 20% validation / 20% test (stratified by type)
• SMOTE-style oversampling applied to minority classes (minimum 500 per type)

Class Distribution

Limitations

• Trained on self-reported MBTI labels which may contain ~30% noise (per MBTIBench, Li et al. 2025)
• Severe class imbalance (47x ratio between INFP and ESTJ) addressed with oversampling but not fully resolved
• Best used as part of the full P2P pipeline with RAG, not as a standalone classifier
• MBTI as a personality framework has limited scientific validity compared to Big Five

Citation

@inproceedings{ma2025p2p,
  title={From Post To Personality: Harnessing LLMs for MBTI Prediction in Social Media},
  author={Ma, Tian and Feng, Kaiyu and Rong, Yu and Zhao, Kangfei},
  booktitle={Proceedings of the 34th ACM International Conference on Information and Knowledge Management},
  year={2025},
  publisher={ACM}
}

Framework Versions

• PEFT 0.18.1
• Transformers 4.57.6
• Unsloth 2026.4.4
• PyTorch 2.10.0