tau-dev-task-airline-v1

Multi-turn tool-calling SFT dataset (268 records, 200 / 18 / 50 train / validation / test) derived from sierra-research/tau-bench airline-domain trajectories.

Meant to be used as a benchmark dataset for developing and validating data processing, training, and eval workflows involving tool use. Note: tau-bench is a widely-used public benchmark and many recently-trained models may have encountered variants of these trajectories during training, so be mindful of potential contamination when interpreting eval results.

Schema

Each record is one conversation in OpenAI tool-calling wire format:

{
  "messages": [{"role": "...", "content": "...", "tool_calls": [...], "tool_call_id": "..."}],
  "tools":    [{"type": "function", "function": {"name": "...", "description": "...", "parameters": {...}}}],
  "metadata": {"record_id": "airline-gpt4o-6-0", "reward": 1.0, "source": "gpt4o", "split": "train", ...}
}

Notes:

• tool_calls[*].function.arguments is a JSON-encoded string, not an object.
• Assistant turns may have content: null when they emit only a tool call (canonical OpenAI shape).
• The tau-bench think tool has been inlined: its thought argument now appears as a <thinking>...</thinking> block in the surrounding assistant content (rather than as a no-op tool call returning ""). think is therefore not in the tools list. The marker <thinking> is intended as regular content — do not confuse with <think>...</think> (a special reasoning token in Qwen3 / DeepSeek-R1 chat templates).
• ###STOP### user-side end-of-conversation sentinels have been stripped.

Provenance

Sources merged: GPT-4o and Claude 3.5 Sonnet trajectories from tau-bench's historical_trajectories/, filtered to reward == 1.0. Random splits via random.Random(666). Upstream tau-bench is MIT licensed.

Upstream now recommends tau2-bench (fixed task quality + additional domains). See lefft/tau-dev-task-airline-v2 (forthcoming) for the refreshed version.