DistilBERT Sentiment Classifier (Run-2 Champion)

Fine-tuned distilbert-base-uncased for 3-class sentiment classification (negative, neutral, positive) on the English subset of the cardiffnlp/tweet_sentiment_multilingual dataset.

Trained with hardware optimizations designed for a 4GB VRAM NVIDIA GeForce RTX 3050 Laptop GPU, including FP16 Tensor Core mixed precision and CUDA-fused AdamW (adamw_torch_fused).

Benchmark Results (Test Set: 870 Unseen Samples)

Metric	Score	Notes
Accuracy	66.90%	+1.04% over 3-epoch baseline
Macro F1	66.51%	Balanced across all 3 classes
Weighted F1	66.51%	Metric used for checkpoint selection
Negative F1 / Recall	72.98% / 82.41%	Robust detection of complaints & negative sentiment
Positive F1 / Precision	73.70% / 79.60%	High precision when flagging positive sentiment
Neutral F1 / Recall	52.84% / 49.66%	Weakest class; neutral is often confused with its neighbors
Severe Polarity Inversion	3.56%	Low confusion between opposing sentiments (31/870 samples)

Confusion Matrix

Training Configuration & Hyperparameters

Base Architecture: distilbert-base-uncased (66.95M parameters)
Epochs: 10 (Champion run with best model selection)
Learning Rate: 2.0e-5 (Linear warmup & decay)
Per-Device Batch Size: 16
Gradient Accumulation Steps: 2 (Effective Batch Size = 32)
Precision: Mixed Precision FP16 (AMP)
Optimizer: adamw_torch_fused (Weight Decay = 0.01)
Hardware Footprint: Peak reserved VRAM of 1,456.0 MB (~1.42 GB / 36.4% utilization)

Quickstart & Usage

Option 1: Hugging Face Pipeline (Simplest)

from transformers import pipeline

# Load pipeline directly from Hugging Face Hub
classifier = pipeline("sentiment-analysis", model="ZyroGod/distilbert-sentiment-classifier")

# Run inference
sample_text = "I absolutely love this product! Best purchase ever!"
prediction = classifier(sample_text)
print(prediction)
# Output: [{'label': 'positive', 'score': 0.7588}]

Option 2: PyTorch AutoModel & AutoTokenizer

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = "ZyroGod/distilbert-sentiment-classifier"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
model.eval()

text = "Amazing customer service and quick resolution!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)[0]

id2label = model.config.id2label
for idx, prob in enumerate(probs):
    print(f"{id2label[idx]:<10s}: {prob.item():.2%}")

Label Mapping

Label ID	Sentiment Class
`0`	`negative`
`1`	`neutral`
`2`	`positive`

License

This model is open-sourced under the Apache 2.0 License.

Loading…

DistilBERT Sentiment Classifier (Run-2 Champion)

Fine-tuned distilbert-base-uncased for 3-class sentiment classification (negative, neutral, positive) on the English subset of the cardiffnlp/tweet_sentiment_multilingual dataset.

Trained with hardware optimizations designed for a 4GB VRAM NVIDIA GeForce RTX 3050 Laptop GPU, including FP16 Tensor Core mixed precision and CUDA-fused AdamW (adamw_torch_fused).

Benchmark Results (Test Set: 870 Unseen Samples)

Metric	Score	Notes
Accuracy	66.90%	+1.04% over 3-epoch baseline
Macro F1	66.51%	Balanced across all 3 classes
Weighted F1	66.51%	Metric used for checkpoint selection
Negative F1 / Recall	72.98% / 82.41%	Robust detection of complaints & negative sentiment
Positive F1 / Precision	73.70% / 79.60%	High precision when flagging positive sentiment
Neutral F1 / Recall	52.84% / 49.66%	Weakest class; neutral is often confused with its neighbors
Severe Polarity Inversion	3.56%	Low confusion between opposing sentiments (31/870 samples)

Confusion Matrix

Training Configuration & Hyperparameters

Base Architecture: distilbert-base-uncased (66.95M parameters)

Epochs: 10 (Champion run with best model selection)

Learning Rate: 2.0e-5 (Linear warmup & decay)

Per-Device Batch Size: 16

Gradient Accumulation Steps: 2 (Effective Batch Size = 32)

Precision: Mixed Precision FP16 (AMP)

Optimizer: adamw_torch_fused (Weight Decay = 0.01)

Hardware Footprint: Peak reserved VRAM of 1,456.0 MB (~1.42 GB / 36.4% utilization)

Quickstart & Usage

Option 1: Hugging Face Pipeline (Simplest)

from transformers import pipeline # Load pipeline directly from Hugging Face Hub classifier = pipeline("sentiment-analysis", model="ZyroGod/distilbert-sentiment-classifier") # Run inference sample_text = "I absolutely love this product! Best purchase ever!" prediction = classifier(sample_text) print(prediction) # Output: [{'label': 'positive', 'score': 0.7588}]

Option 2: PyTorch AutoModel & AutoTokenizer

import torch from transformers import AutoTokenizer, AutoModelForSequenceClassification model_name = "ZyroGod/distilbert-sentiment-classifier" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained(model_name) model.eval() text = "Amazing customer service and quick resolution!" inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128) with torch.no_grad(): outputs = model(**inputs) probs = torch.softmax(outputs.logits, dim=-1)[0] id2label = model.config.id2label for idx, prob in enumerate(probs): print(f"{id2label[idx]:<10s}: {prob.item():.2%}")

Label ID

Sentiment Class

0

negative

1

neutral

2

positive