| --- |
| language: en |
| license: mit |
| library_name: transformers |
| tags: |
| - sentiment-analysis |
| - classification |
| - from-scratch |
| - multi-domain |
| datasets: |
| - imdb |
| - glue |
| metrics: |
| - accuracy |
| model-index: |
| - name: VibeCheck-v1 |
| results: |
| - task: |
| type: text-classification |
| name: Sentiment Analysis |
| dataset: |
| name: Mixed (IMDb & SST2) |
| type: multi-domain |
| metrics: |
| - type: accuracy |
| value: [INSERT_ACCURACY_HERE] |
| pipeline_tag: text-classification |
| --- |
| |
| # VibeCheck v1 |
|
|
| VibeCheck v1 is a high-performance, multi-domain Transformer model trained **entirely from scratch**. Unlike its predecessor, this model was trained on a balanced mix of long-form reviews and short-form conversational data, making it a versatile tool for analyzing "vibes" across different types of English text. |
|
|
| ## Model Description |
| - **Architecture:** Enhanced Custom Transformer (DistilBERT-style) |
| - **Parameters:** ~11.17 Million |
| - **Layers:** 4 (Increased depth for better abstraction) |
| - **Attention Heads:** 8 |
| - **Hidden Dimension:** 256 (Hidden Feed-Forward: 1024) |
| - **Training Data:** ~92,349 samples (Mixed IMDb Movie Reviews & SST-2 Sentence Bank) |
| - **Training Duration:** ~25-30 minutes on NVIDIA T4 GPU |
|
|
| ## Capabilities |
| - **Multi-Domain Versatility:** Reliable on everything from formal emails to short chat messages. |
| - **Enhanced Context Awareness:** 4 layers of self-attention allow for a deeper understanding of sentence structure. |
| - **Linguistic Nuance:** Strong performance on complex negatives (e.g., "not as bad as I thought") and rhetorical questions. |
| - **Robustness:** High tolerance for slang, typos, and non-standard English. |
|
|
| ## Limitations |
| - **Language Focus:** Primarily trained on English. While it shows some intuition for other languages, accuracy may vary. |
| - **Binary Nature:** Strictly classifies text as Positive or Negative; it does not detect neutral intent or specific emotions (like anger or joy). |
|
|
| ## How to use (Inference Script) |
| To use this model, download the `VibeCheck_v1_Model.zip`, unpack it, and run the provided `inference.py` script. Make sure to point the script to the unpacked directory. |
|
|
| ## Examples (VibeCheck v1 in Action) |
|
|
| ### Example 1: Formal Business Email |
| **Input:** |
| ```plaintext |
| Dear Team, I am writing to express my deep disappointment regarding the recent project update. The quality is subpar. |
| ``` |
|
|
| **Output:** NEGATIVE | Confidence: 97.60% |
|
|
| ### Example 2: Short Conversational Fragment |
| **Input:** |
|
|
| ```plaintext |
| That sounds like a fantastic plan! I'm starving. |
| ``` |
|
|
| **Output:** POSITIVE | Confidence: 74.15% |
|
|
| ### Example 3: Sarcastic Observation of a movie review |
| **Input:** |
|
|
| ```plaintext |
| Wow! What an amazing view we have out of this window! |
| ``` |
|
|
| **Output:** POSITIVE | Confidence: 99.43% |
|
|
| ### Example 4: Classic Test |
| **Input:** |
|
|
| ```plaintext |
| You are dumb. |
| ``` |
|
|
| **Output:** NEGATIVE | Confidence: 87.85% |
|
|
| ### Example 5: Simple chat |
| **Input:** |
|
|
| ```plaintext |
| Did you see the new movie?' B: 'Yeah, it was okay, but the ending felt a bit rushed.' A: 'I totally agree, it could have been better.' |
| ``` |
|
|
| **Output:** NEGATIVE | Confidence: 80.98% |
|
|
| ## Training code |
| The full training code for this multi-domain version is available in `train.ipynb`. |
