data mixture

**Pretraining Data Mixture and Curation** is the **strategic selection and weighting of training data domains that critically determines the capabilities, biases, and performance characteristics of large language models** — where the composition of web text, books, code, scientific papers, dialogue, and multilingual content in the training mixture has a larger impact on model quality than architecture differences, making data curation one of the most important and closely guarded aspects of frontier LLM development. **Why Data Mixture Matters** - Same architecture + same compute + different data mixture → dramatically different models. - Code data improves reasoning (even for non-code tasks). - Math data enables quantitative reasoning. - Book data improves long-range coherence. - Web data provides breadth but includes noise. **Data Source Characteristics** | Source | Volume | Quality | What It Teaches | |--------|--------|---------|----------------| | Common Crawl (web) | 100T+ tokens | Low-medium | Breadth, world knowledge | | Wikipedia | ~4B tokens | High | Factual knowledge, structure | | Books (BookCorpus, etc.) | ~5B tokens | High | Long-form coherence, reasoning | | GitHub/StackOverflow | ~100B tokens | Medium-high | Code, structured thinking | | ArXiv/PubMed | ~30B tokens | High | Scientific reasoning | | Reddit/forums | ~50B tokens | Medium | Dialogue, opinions | | Curated instruction data | ~1B tokens | Very high | Task following | **Known Model Mixtures** | Model | Web | Code | Books | Wiki | Other | |-------|-----|------|-------|------|-------| | Llama 1 | 67% | 4.5% | 4.5% | 4.5% | 19.5% (CC-cleaned) | | Llama 2 | ~80% | ~10% | ~4% | ~3% | ~3% | | Llama 3 | ~50% | ~25% | ~10% | ~5% | ~10% | | GPT-3 | 60% | 0% | 16% | 3% | 21% | | Phi-1.5 | 0% | 0% | 0% | 0% | 100% synthetic | **Data Filtering Pipeline** ``` [Raw Common Crawl: ~300TB compressed] ↓ [Language identification] → Keep target languages ↓ [URL and domain filtering] → Remove known low-quality sites ↓ [Deduplication] → MinHash + exact dedup → removes 40-60% ↓ [Quality classifier] → FastText trained on curated vs. random → remove bottom 50% ↓ [Content filtering] → Remove toxic, PII, CSAM ↓ [Domain classification] → Tag and weight by domain ↓ [Final mixture: ~5-15T high-quality tokens] ``` **Data Mixing Strategies** | Strategy | Approach | Used By | |----------|---------|--------| | Proportional | Sample proportional to domain size | Early models | | Upsampled quality | Oversample high-quality domains (Wikipedia, books) | GPT-3, Llama 1 | | DoReMi | Optimize domain weights via proxy model | Google | | Data mixing laws | Predict performance from mixture via scaling laws | Research frontier | | Curriculum | Start with easy/clean data, add harder data later | Some proprietary models | **Deduplication Impact** - Training on duplicated data: Memorization increases, generalization decreases. - Exact dedup: Remove identical documents → easy, removes ~20%. - Near-dedup (MinHash): Remove ~similar documents → removes additional 20-40%. - Effect: Deduplication equivalent to 2-3× more unique training data. **Data Quality vs. Quantity** | Approach | Data | Model | Result | |----------|------|-------|--------| | Llama 2 (70B) | 2T tokens (web-heavy) | 70B | Strong general | | Phi-2 (2.7B) | 1.4T tokens (curated + synthetic) | 2.7B | ≈ Llama 2 7B quality | | FineWeb-Edu | Web filtered for educational content | Various | Significant improvement | Pretraining data curation is **the most impactful yet least understood lever in LLM development** — while architectural innovations yield marginal gains, the choice of which data to train on and in what proportions fundamentally determines a model's capabilities, with frontier labs investing millions of dollars and years of effort into data pipelines that are among their most carefully protected competitive advantages.

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account