| --- |
| license: mit |
| task_categories: |
| - other |
| tags: |
| - tool-calling |
| - agents |
| - synthetic |
| - travel |
| pretty_name: ToolSmith Tasks |
| --- |
| |
| # ToolSmith Tasks |
|
|
| Synthetic travel-ops tool-calling tasks for training and evaluating Qwen3-4B-Instruct-2507 |
| (via LoRA SFT + step-level GRPO), generated against the ToolSmith deterministic 12-tool sandbox. |
|
|
| ## Dataset Summary |
|
|
| - **Tasks:** 594 across 4 difficulty tiers (T1 single-tool, T2 2-tool chains, T3 4-6 tool |
| chains with dependencies, T4 traps — unsolvable/ambiguous requests where the correct |
| behavior is to decline or ask for clarification, not hallucinate a tool call) |
| - **Splits:** train / val / test, stratified by tier (417 / 89 / 88) |
| - **Generation:** synthetic, produced by a local template-based generator |
| (`scripts/generate_tasks_local.py`) parameterized across the sandbox's real world data |
| (`src/toolsmith/tools/sandbox/worlddata/`) — an alternative to prompting Claude directly |
| (`scripts/generate_tasks.py`) for environments without a live Anthropic API key. Every |
| multi-step task's dependent values (e.g. a city's lat/lon before a weather lookup) are |
| computed by actually executing the relevant sandbox tool, not invented. Validated for 100% |
| solvability by a bounded BFS solver (`src/toolsmith/data/solver.py`); T4 traps are trivially |
| "solvable" with zero tool calls by design (the correct answer never needs one). |
| - **Goal specs:** every task carries a machine-checkable goal spec (not a gold trajectory) — |
| rewards verify sandbox outcomes, not paths |
|
|
| ## Fields |
|
|
| | Field | Type | Description | |
| |---|---|---| |
| | `id` | string | unique task id | |
| | `tier` | string | one of T1, T2, T3, T4 | |
| | `user_prompt` | string | the natural-language traveler request | |
| | `goal_spec` | list | machine-checkable conditions (see below) | |
| | `min_steps` | int | solver-computed minimum sandbox tool calls to satisfy the goal | |
| | `split` | string | one of train, val, test | |
|
|
| ### Goal condition types |
|
|
| - `answer_contains_fact` — final answer text must contain a given substring |
| - `tool_was_called_with` — a successful tool call must match a tool name + arg subset |
| - `calendar_event_exists` — a `calendar_create_event` call with exact fields must have succeeded |
| - `numeric_within_tolerance` — a number from the final answer or a tool result must be near |
| an expected value |
|
|
| ## How This Dataset Is Used |
|
|
| - **SFT** (`notebooks/src/01_sft_warmstart.py`): `train`-split tasks are replayed through an |
| agent inside the episode runner; trajectories whose goal spec passes become gold SFT rows |
| (`scripts/generate_gold_trajectories.py`, or its local equivalent |
| `scripts/generate_gold_trajectories_local.py`, which scripts the exact required tool-call |
| sequence from each goal_spec directly rather than improvising it, since every T1-T3 |
| goal_spec already IS that sequence). |
| - **GRPO** (`notebooks/src/02_grpo_training.py`): `goal_spec` feeds the R5 outcome reward |
| (`src/toolsmith/rewards/outcome_reward.py`) directly — every candidate action is scored by |
| executing it in the sandbox and checking these same conditions. `min_steps` feeds the R6 |
| efficiency bonus. |
| - **Eval** (`src/toolsmith/eval/runner.py`): the `test` split is the held-out 4-way comparison |
| suite (88 tasks); see the model card's evaluation table for results. |
|
|
| ## Contamination Controls |
|
|
| SFT gold trajectories are drawn only from the `train` split; `test` never touches training. |
| This dataset was generated independently of, and shares no rows with, the xLAM/Glaive public |
| function-calling corpora used for SFT warm-start. |
|
|
| ## License |
|
|
| MIT. |
|
|
