aimosprite/brian-rollouts-311-training-turns

# brian-rollouts-311-training-turns-v1 Lean per-turn GPT-OSS training export derived from `brian-rollouts-311-full-v2`. Contents: - `rollouts.training_turns.jsonl`: one row per assistant call with full prompt token IDs and completion token IDs - `manifest.json`: export metadata and row counts Semantics: - `prompt_token_ids` are the full context shown to the model for that assistant call. - `completion_token_ids` are the tokens generated by the model on that call. - `prompt_token_ids` include system prompt, problem text, prior assistant turns, and tool-result user messages. - `completion_token_ids` exclude invisible terminal tokens such as `<|return|>` because they are not recoverable from saved text.