| --- |
| license: cc-by-4.0 |
| task_categories: |
| - text-generation |
| tags: |
| - agentic |
| - tool-use |
| - function-calling |
| - multi-turn |
| - error-recovery |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # Agentic Tool-Use Recovery (SFT) |
|
|
| Synthetic multi-turn tool-use trajectories that teach **recovery from failed tool results**, **query reformulation**, and **when to go direct vs. recover**. Built to address the failure mode seen when a Qwen2.5-7B model SFT'd on APIGen-MT (happy-path only) was evaluated on TAC: it called tools correctly but, when a search returned empty, it re-issued near-identical queries in a loop instead of adapting, and rarely completed the task. |
|
|
| Each trajectory is a travel-booking agent that must `search_experiences` → `get_experience_details` → `check_availability` → `purchase_tickets`. |
|
|
| **Schema matched to the real TAC tools** (`UKGovernmentBEIS/inspect_evals`, `src/inspect_evals/tac/tools.py`): search's optional filter parameter is `keywords` (not `query`), `check_availability` takes an optional `num_tickets`, results include Tags, and the success/failure return strings mirror TAC's wording, so the model trains on the same observations it will see at eval time. TAC matches `location` by substring of any term, which is why an over-broad term ("Hawaii") fails against an experience filed under "Honolulu, HI" while the city name works. The trajectories reflect that. |
|
|
| ## Composition — 3,750 trajectories |
|
|
| Recovery (3,000, ~25% each): |
| - **broad_to_city** — first search uses a too-broad location → empty → narrow to the city → results → book. |
| - **two_wrong_then_city** — two over-broad/wrong location strings fail before narrowing to the city. |
| - **keyword_refine** — city search returns many → use the `keywords` field to narrow to the target. |
| - **bad_keyword_recover** — an over-narrow/garbled `keywords` returns empty → broaden the keyword → results → book. |
|
|
| Direct success (750, ~20% of the set): |
| - **direct_success** — the user names a city, the agent searches it directly (sometimes with `keywords`) and books, no failure. This teaches the *discrimination* (go direct when you already have a city) so the model doesn't learn a "fail-first ritual" from an all-recovery set. |
| |
| ## Schema |
| |
| JSONL, one trajectory per line: |
| |
| ```json |
| { |
| "messages": [ |
| {"role": "system", "content": "..."}, |
| {"role": "user", "content": "..."}, |
| {"role": "assistant", "content": "reasoning...", "tool_calls": [{"id": "...", "type": "function", "function": {"name": "search_experiences", "arguments": {"location": "Hawaii"}}}]}, |
| {"role": "tool", "tool_call_id": "...", "name": "search_experiences", "content": "No experiences found in 'Hawaii'. Try a different location or broader keywords."}, |
| {"role": "assistant", "content": "...", "tool_calls": [{"...": "search Honolulu ..."}]}, |
| {"role": "tool", "...": "Found N experience(s) in Honolulu ..."}, |
| "... details -> availability -> purchase ...", |
| {"role": "assistant", "content": "All set! I booked ..."} |
| ], |
| "tools": [ /* OpenAI-style function schemas, matched to TAC */ ], |
| "metadata": {"city": "...", "country": "...", "target_id": "...", "category": "...", "pattern": "...", "num_tickets": 1} |
| } |
| ``` |
| |
| - `tool_calls[].function.arguments` is a **dict** (Qwen2.5 `apply_chat_template` friendly; `json.dumps` it if your trainer wants a string). |
| - Apply your chat template with `tools=row["tools"]`; mask loss to assistant turns only. |
|
|
| ## Intended use |
|
|
| Mix as roughly **40%** of a small tool-use SFT set, with ~60% APIGen-MT (multi-turn backbone), ~3 epochs. This set teaches recovery + the direct/recover discrimination; it is meant to be combined, not used alone. |
|
|
| ## Caveats |
|
|
| - **Synthetic / templated.** Realistic in structure, not scraped; phrasing variety is from template pools. |
| - **Benign bookings only.** Options are welfare-neutral (hiking, snorkeling reefs, cooking classes, etc.). Targets completion/recovery, not welfare selection; do not expect it to move a welfare metric. |
| - **Not a benchmark.** Training data only; no held-out eval. Validate the effect on TAC `completion_rate` directly, do not assume. |
|
|
| Generated 2026-07-02, deterministic seed. See `gen_agentic_recovery.py`. |
|
|
