SCHEMA.md

Schema (SOTA rc2 subsets)

Base structured configs reuse compile/materialize column families:

• sft* / openai_messages: parent_id, id, split, messages, tools, sampling_weight, source, domain
• canonical / smoke: full audit columns
• rl_tool_prompts: family, prompt_messages_json, tools_json, expected_terminal_oracle_json, reward_contract_json
• glm47_native: text, renderer_revision (chat-template text; not input_ids tensors in rc2)
• token_stats: assistant_token_estimate, context_bucket (estimate; exact retokenization optional follow-up)