metadata
license: apache-2.0
pipeline_tag: text-classification
tags:
- sentiment-analysis
- text-classification
- from-scratch
- transformer
- sst2
datasets:
- stanfordnlp/sst2
metrics:
- accuracy
Vibe Check — a sentiment classifier trained from scratch
Vibe Check is a compact (~2M parameter) Transformer text classifier that predicts whether a sentence is Positive or Negative. It was built entirely from scratch — my own word tokenizer and my own model architecture, trained from random initialization (no pretrained weights, no fine-tuning) on the public Stanford Sentiment Treebank (SST-2) dataset.
Highlights
- From scratch: custom tokenizer + 2-layer Transformer encoder + mean-pooling classifier head.
- ~2.0M parameters, trained from random init on ~67k public sentences.
- ~82% validation accuracy on SST-2 (small model, no pretrained embeddings).
- Runs on CPU in milliseconds.
Not a fine-tune
Unlike many models on the Hub, this is not a fine-tune of a large pretrained model — every weight was learned from zero on public data. Training code, tokenizer, and config are included.
Usage
# see vibe.py in this repo for the model definition + predict_vibe()
from vibe import predict_vibe
print(predict_vibe("I absolutely love this!")) # {'label': 'Positive', 'confidence': 98.5}
Data & license
Trained on SST-2 (public research dataset). Released under Apache-2.0. For entertainment/educational use; a small model, so expect ~1 in 5 predictions to be wrong on hard/ambiguous text.