# Dataset attribution

The Coder-Instruct v1 SFT corpus is deterministic project-authored material
licensed under Apache-2.0. It imports no external instruction dataset, private
data, chat exports, Telegram data, or chain-of-thought material.

## Composition

- Train: 2,640 examples, 323,960 rendered tokens, 88 source groups.
- Validation: 330 examples, 40,504 rendered tokens, 11 source groups.
- Test: 330 examples, 40,518 rendered tokens, 11 source groups.
- Maximum rendered sequence: 149 tokens; model limit: 1,024.
- Eleven balanced task types: short function, completion, explanation, bug fix,
  refactor, unit test, traceback, JSON/YAML conversion, SQL, PowerShell, and FIM repair.
- Source groups are disjoint across train, validation, and test.

## Gates

License/provenance, secret, email-like PII, exact deduplication, cross-split
source-group leakage, chain-of-thought exclusion, and chat-marker collision
gates passed. Bounded local validators passed for Python AST/fragments, JSON,
YAML, SQL, safe PowerShell, and concise text.

This conservative, template-heavy corpus validates the SFT pipeline and balanced
held-out evaluation. It does not establish broad coding-assistant competence.