Natural language processing - Analyzing and Mitigating Dataset Artifacts
Budget: ₹1,500 – ₹12,500 INR
I will share a short PDF brief and a worked example that outline a small-scale experiment aimed at analysing and mitigating dataset artifacts in natural-language-processing workflows. Your role is to reproduce the study in Python, refine the methods where helpful, and document the whole process in a concise 3-to-7-page report.
The core of the assignment is to pinpoint how hidden artifacts influence a model’s performance, demonstrate mitigation techniques, and show the before-and-after metrics. Although the PDF leaves model choice open, a lightweight text-classification pipeline built with PyTorch, TensorFlow or scikit-learn will be enough—the emphasis is on the artifact investigation rather than SOTA accuracy.
Key deliverables
• Clean, well-commented Python code (script or notebook) that loads the data, surfaces the artifacts, applies your mitigation strategy, and reruns evaluation.
• A 3–7 page write-up (PDF, Word, or Markdown) that covers methodology, findings, visualisations, and a brief discussion of limitations and future work.
Acceptance criteria
1. The report clearly contrasts model performance before vs. after artifact mitigation.
2. Code runs end-to-end with instructions and environment details.
3. All figures in the report are generated by the supplied code.
The scope is intentionally compact—ideal for a focused weekend project—yet demands sound machine-learning judgement and clear communication.
COde repo:
https://github.com/gregdurrett/fp-dataset-artifacts
The core of the assignment is to pinpoint how hidden artifacts influence a model’s performance, demonstrate mitigation techniques, and show the before-and-after metrics. Although the PDF leaves model choice open, a lightweight text-classification pipeline built with PyTorch, TensorFlow or scikit-learn will be enough—the emphasis is on the artifact investigation rather than SOTA accuracy.
Key deliverables
• Clean, well-commented Python code (script or notebook) that loads the data, surfaces the artifacts, applies your mitigation strategy, and reruns evaluation.
• A 3–7 page write-up (PDF, Word, or Markdown) that covers methodology, findings, visualisations, and a brief discussion of limitations and future work.
Acceptance criteria
1. The report clearly contrasts model performance before vs. after artifact mitigation.
2. Code runs end-to-end with instructions and environment details.
3. All figures in the report are generated by the supplied code.
The scope is intentionally compact—ideal for a focused weekend project—yet demands sound machine-learning judgement and clear communication.
COde repo:
https://github.com/gregdurrett/fp-dataset-artifacts