somaxsoma/tac-closing-efficiency-sft
TAC closing-efficiency slice 500 synthetic multi-turn tool-use trajectories that teach an agent to close bookings decisively — the welfare-neutral capability piece of the tool-use SFT mix used to train somaxsoma/qwen2.5-7b-tac-recovery-sft. What it teaches Built to fix the dominant failure mode observed on the TAC benchmark — the model reformulating search keywords in a loop and never closing a booking. Three patterns: settle/browse (200): after failed keyword… See the full description on the dataset page: https://huggingface.co/datasets/somaxsoma/tac-closing-efficiency-sft.
TAC closing-efficiency slice
500 synthetic multi-turn tool-use trajectories that teach an agent to close bookings decisively — the welfare-neutral capability piece of the tool-use SFT mix used to train `somaxsoma/qwen2.5-7b-tac-recovery-sft`.
What it teaches
Built to fix the dominant failure mode observed on the TAC benchmark — the model reformulating search keywords in a loop and never closing a booking. Three patterns:
- settle/browse (200): after failed keyword searches, drop the keywords, browse the location, and settle for the closest available option, then complete the booking.
- direct-book (200): once any results appear, stop searching and proceed details → availability → purchase.
- availability pivot (100): when the first choice is sold out, pivot to an alternative from the same result set rather than re-searching.
Welfare-neutral by construction
The generator (build_efficiency_slice.py) asserts no animal/wildlife/animal-venue content appears in any message, so this slice teaches decisive closing without injecting welfare signal. (In the research design, welfare is instilled via mid-training, not this capability SFT.)
Format
One JSON object per line: {"messages": [...], "tools": [...], "source": "efficiency"} — unified chat schema (role-tagged messages + OpenAI-style tool list; tool-call arguments as dicts), matching the Qwen2.5 chat template.
Full training mix (this is one of three pieces)
The model was trained on this slice plus:
- `Salesforce/APIGen-MT-5k` — multi-turn tool-use backbone (gated; accept terms on its page)
- `CompassioninMachineLearning/agentic-tool-recovery-sft` — recovery from failed/empty tool calls
Reproduce the exact combined set with build_dataset.py (see the project repo). APIGen is not re-hosted here due to its gating.
Generation
Deterministic (seed 7): python build_efficiency_slice.py. Generator and full training/eval pipeline in the project repo.