---
library_name: transformers
license: mit
base_model: xlm-roberta-base
tags:
- generated_from_trainer
- mbti
- midwest-emo
- math-rock
- personality-detection
- domain-adaptation
- hybrid-corpus
metrics:
- accuracy
datasets:
- anggars/mbti-emotion
language:
- id
- en
model-index:
- name: xlm-mbti
  results:
  - task:
      type: text-classification
      name: Text Classification
    dataset:
      name: anggars/mbti-emotion
      type: anggars/mbti-emotion
    metrics:
    - type: accuracy
      value: 0.4433
      name: Accuracy
---

# XLM-RoBERTa MBTI (Domain-Adapted for Midwest Emo/Math Rock)

This model is a fine-tuned version of [xlm-roberta-base](https://huggingface.co/xlm-roberta-base) for **MBTI Personality Classification** (16 types). It has been architecturally recalibrated using a **Hybrid Corpus** to extract underlying cognitive functions (*Thinking, Feeling, Intuition, Sensing*) from poetic hyperboles and complex metaphors, specifically within the context of Midwest Emo and Math Rock lyrics.

## Model Description

- **Model Type:** XLM-RoBERTa Base (Sequence Classification Head with 16 Nodes)
- **Labels:** 16 MBTI Personality Types
- **Dataset:** `anggars/mbti-emotion` (Hybrid Corpus: 112,351 synthetic narrative rows + organic scraped lyrics)
- **Language(s):** English & Indonesian (Multilingual)
- **License:** MIT
- **Training Environment:** Kaggle Compute (Dual Tesla T4 GPU, fp16 Mixed Precision)

## Architectural Innovations & Trade-offs

Predicting 16 distinct personality classes purely from unstructured text is a highly complex NLP task (random guessing yields only a 6.25% baseline). 

In this iteration, the model underwent **Domain Adaptation** via a Hybrid Corpus to eliminate its dependency on rigid, formal synthetic narratives. By forcing the architecture to learn from actual organic lyrics scraped from Genius.com, the model developed zero-shot capabilities for real-world musical analysis.

*Note on Metrics: The shift in Validation Loss (to ~3.3) and Accuracy (to ~44.3%) compared to earlier synthetic-only baselines is a deliberate architectural trade-off. The implementation of aggressive weight decay (0.05) acts as a natural **Label Smoothing** mechanism. Since organic lyrics rarely belong 100% to a single MBTI profile (e.g., a lyric might possess traits of both INFP and ISFP), the model distributes its probability confidence via **Softmax Calibration**. This prevents overconfident hallucination and yields highly organic, generalized predictions suitable for production environments.*

## Training Results

The following results were achieved on the evaluation set during the 3-epoch background training process:

| Epoch | Step  | Validation Loss | Accuracy |
|:-----:|:-----:|:---------------:|:--------:|
| 1.0   | 5624  | 3.5125          | 0.4006   |
| 2.0   | 11248 | 3.3358          | 0.4341   |
| 3.0   | 16872 | 3.3074          | 0.4433   |

## Intended Uses & Limitations

This model is intended for academic research in the field of Natural Language Processing (NLP) and psychology, specifically functioning as the backend engine for music analytics dashboards. 
**Limitations:** Personality cognitive functions are highly complex. The model provides predictions based strictly on linguistic and lyrical patterns in specific musical subgenres. It operates on poetic heuristics and must not be utilized as a definitive psychological diagnostic tool for human subjects.

## Training Procedure

### Training Hyperparameters

To ensure stable convergence on metaphorical organic lyrics and prevent catastrophic forgetting of the synthetic distribution, the following hyperparameters were enforced:

- **learning_rate:** 1.5e-05
- **train_batch_size:** 16
- **eval_batch_size:** 16
- **seed:** 42
- **weight_decay:** 0.05
- **optimizer:** AdamW with betas=(0.9,0.999) and epsilon=1e-08
- **lr_scheduler_type:** linear
- **num_epochs:** 3
- **mixed_precision_training:** Native AMP (fp16)

### Framework Versions

- Transformers 4.44.2
- Pytorch 2.5.1+cu124
- Datasets 3.1.0
- Tokenizers 0.20.3