gradients-io-tournaments/pvp-tool-calling-sft

PvP tool-calling SFT cold-start data

Claude-vs-Claude games played through the G.O.D PvP tool-calling harness. Each row is one model turn (or post-game reflection): the system+user prompt the harness built, the assistant response (content + tool_calls), and the tools schemas — i.e. the OpenAI messages+tools format consumed by tokenizer.apply_chat_template(messages, tools=tools). On a move turn the assistant co-emits any memory-tool edits and a game_action committing a legal move; reflection rows consolidate long-term memory after a game.

Players: claude-sonnet-4-6 vs claude-haiku-4-5 (positions swapped each seed, so both orderings appear).

Total: 850 samples across 5 environments.

Per-env sample counts

• —gin_rummy: 170
• —goofspiel: 170
• —leduc_poker: 170
• —liars_dice: 170
• —othello: 170

Fields

• —messages, tools: the training example (OpenAI chat + tools format)
• —env: game name
• —model: the Claude model that authored the sample
• —turn_type: turn (commits game_action) or reflection (consolidates long-term memory)
• —seq: chronological play order within the env file (stable across shuffling)
• —arc: unique long-term-memory arc id — memory persists across the games within one arc
• —turn_in_arc: order of the sample within its arc
• —matchup: per-generation-run arc index (legacy; use arc for a globally-unique id)

To reconstruct a long-term-memory trajectory, take one (env, arc) group and sort by turn_in_arc (or seq). For plain SFT this is unnecessary — each row is self-contained, since the memory state at that turn is already rendered inside its prompt.