Add tokenizer docs

tokenizer/PREPROCESSING.md

@@ -0,0 +1,22 @@
+# Tokenizer Preprocessing
+
+Tokenizer training should sample across files instead of consuming the first
+large file only. Use round-robin sampling from checkpoint shards to avoid
+language and domain skew.
+
+## Keep
+
+- Exact indentation.
+- Blank lines.
+- Comments and docstrings.
+- Mixed Korean/English comments.
+- Repository/file/language metadata when represented in text.
+
+## Drop Or Quarantine
+
+- Empty text.
+- Null bytes.
+- Obvious credentials, tokens, private keys, and secrets.
+- Malformed JSONL.
+- Extremely long records that exceed the active model context unless the loader
+  has deterministic chunking.

tokenizer/README.md

@@ -0,0 +1,36 @@
+# Tokenizer
+
+## Recommended Tokenizer
+
+Use a byte-level BPE tokenizer with explicit FIM and metadata tokens. The
+current local tokenizer uses Hugging Face `tokenizers` JSON format.
+
+Required special tokens:
+
+- `<|fim_prefix|>`
+- `<|fim_suffix|>`
+- `<|fim_middle|>`
+- `<|fim_pad|>`
+- `<|repo|>`
+- `<|file|>`
+- `<|lang|>`
+- `<|endoftext|>`
+- `<|pad|>`
+
+## Usage Rules
+
+- Encode with `add_special_tokens=False` when the record text already contains
+  FIM markers.
+- Decode audits with `skip_special_tokens=False`.
+- Preserve indentation, tabs, newlines, comments, and Korean text.
+- Do not lowercase or normalize code whitespace.
+- Append EOS between JSONL records during training.
+
+## Quality Checks
+
+Audit tokenizer quality with:
+
+- Special tokens remain atomic.
+- Round-trip decode has no byte loss.
+- FIM markers remain visible in decoded samples.
+- Code chars/token is stable across Python, Rust, C++, JavaScript, and Java.