Document custom mixing and Chinchilla scaling

README.md

@@ -42,3 +42,47 @@ See `catalog.json` for revisions, upstream URLs, document counts, hashes, and sa
 - `text`
 
 Use `source_key` or `category` to build a weighted mixture. `chunk_bytes` is the exact UTF-8 size of the row and can be summed for accounting.
+
+## Build your own mix
+
+Load the table, group by `source_key`, and sample rows with weights. This example gives Wikipedia 40%, EUR-Lex 30%, and literature 30%:
+
+```python
+from datasets import load_dataset
+from collections import defaultdict
+import random
+
+ds = load_dataset("SlayerLab/fabryka-track-polish-mix", split="train", streaming=True)
+weights = {"wikipedia": .40, "eurlex": .30, "wolne_lektury": .15, "wikisource": .15}
+rows = ((r["text"], r["source_key"]) for r in ds if r["source_key"] in weights)
+by_source = defaultdict(list)
+for text, source in rows:
+    by_source[source].append(text)
+mix = []
+for source, weight in weights.items():
+    n = round(1000 * weight)
+    mix.extend(random.choices(by_source[source], k=n))
+random.shuffle(mix)
+```
+
+For a streaming pipeline, keep a per-source iterator and draw the next source with `random.choices(..., weights=...)`; this avoids materializing the full corpus. Use `chunk_bytes` to stop at an exact byte-token budget.
+
+![Corpus size by source](source_sizes.png)
+
+## What can this corpus train?
+
+The canonical expanded sources contain **655.4M UTF-8 byte tokens**. With the current byte-token trainer, a useful planning rule is:
+
+| Model size | 20:1 budget | Fit in this corpus? |
+|---:|---:|:---:|
+| 8M | 160M tokens | Yes |
+| 16M | 320M tokens | Yes |
+| 32M | 640M tokens | Barely; only ~2.4% remains |
+| 50M | 1B tokens | No |
+| 100M | 2B tokens | No |
+
+The 20:1 line is a Chinchilla-style planning heuristic, not a guarantee of model quality. This dataset uses bytes rather than a subword tokenizer, so the equivalent BPE-token count and the best training ratio will differ.
+
+![Chinchilla-style scaling budget](scaling_budget.png)
+
+For a robust 32M run, target at least 800M–1B fresh tokens or allow a documented repeat factor. Larger models require more source text or a larger mix.