Model Card: DistilBERT MLOps Demo (CI/CD Pipeline)
Training Details
Training Procedure
- Experimental Phase: Rapid iteration on Google Colab (T4 GPU).
- Production Phase: Automated training via GitHub Actions runners.
Hyperparameters
- Method: PEFT (LoRA)
- Precision: FP16 Mixed Precision
- Optimizer: AdamW
Evaluation
Results Table
| Run Type | Epochs | Train Samples | Eval Samples | Accuracy | Runtime |
|---|---|---|---|---|---|
| Colab GPU (manual) | 3 | ~5,000 | ~1,000 | 85.3% | ~30 mins |
| GitHub Actions (CI/CD) | 2 | Full Pipeline | ~1,000 | 86.0% | ~5 hours |
Summary:
The automated CI/CD run achieved 86.0% accuracy, proving the pipeline can autonomously produce production-ready models with reproducibility.
Limitations & Risks
- Inherits biases from DistilBERT’s pretraining corpus (Wikipedia + Toronto Book Corpus).
- Restricted to English text.
- Not suitable for sensitive domains (medical, legal, financial) without domain-specific fine-tuning.
Environmental Impact
By using LoRA instead of full fine-tuning, compute requirements were reduced by ~90%, lowering both carbon footprint and operational costs.
Contact: Jeya Prakash I.
Project Goal: Bridging the gap between ML Research and Cloud Production.
- Downloads last month
- 6
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for JeyaPrakashI/distilbert-mlops-demo-ci
Base model
distilbert/distilbert-base-uncased