Large datasets can contain repeated documents, low-quality text, corrupted files, spam and evaluation contamination. A dataset may grow in size while adding very little new information.
Quantity Versus Information
Two datasets with the same number of tokens can differ dramatically in usefulness. Diversity, accuracy, clarity, domain balance and duplication all affect what a model can learn.
| Problem | Possible consequence |
|---|---|
| Exact duplication | Wasted compute and memorization |
| Near duplication | Over-representation of repeated ideas |
| Low-quality text | Weak patterns and unstable style |
| Benchmark contamination | Misleading evaluation |
| Poor domain balance | Uneven capability |
Filtering Is a Research Decision
Filtering is not neutral. A rule designed to remove spam can also remove dialects, informal writing or niche technical material. Every filter expresses assumptions about what counts as valuable data.
Deduplication
Exact deduplication can be handled with hashes. Near-duplicate detection requires similarity methods such as MinHash, locality-sensitive hashing or embedding-based retrieval. The right method depends on the scale and the kinds of repetition present.
My Small-Scale Experiment
I will construct two corpora from the same raw source. The first will receive only basic cleaning. The second will include duplication checks, minimum quality rules and document-level language filtering. Both models will use the same architecture and token budget.
Evaluation Questions
- Does the filtered dataset reduce validation loss faster?
- Does it improve factual consistency?
- Does it reduce repetitive generations?
- Which useful documents are accidentally removed?