rapidchat-data 📦

_{Frontier scores: official τ²-bench leaderboard (user simulator gpt-5.2, standard harness). rapidchat: our runs, user simulator gpt-6.1-sol, custom harness, train split seen in training. Different setups, not a like-for-like ranking.}

everything that went into rapidchat: the practice world, the scenarios, the conversations it learned from, and the receipts.

what's in the box

path	what it is
scenarios/db.json	a re-skinned airline world: 500 users, 2000 reservations, 300 flights. Same schema as the τ²-bench airline database, every person, booking and flight renamed.
scenarios/tasks.json	1717 practice scenarios in τ²-bench task format, each with a simulated-customer script and the gold tool calls
scenarios/split_tasks.json	the chunks the scenarios were run in
sft/	chat-format training examples (system prompt, tool schemas, full conversation incl. tool calls and kept reasoning)
eval/	full τ²-bench airline trajectories for both reported runs (50 tasks × 4 trials each)
code/	the scenario generators, validators, SFT builder, LoRA trainer and merge script
working_notes.txt	the airline-policy reminders used in the "with notes" setup

scenario families

family	scenarios
cancel_ok	213
cancel_denied	177
passenger_update	131
two_requests	116
cancel_flown	101
sweep_cancel	97
compensation_false_claim	84
sweep_cabin	78
baggage_add	74
booking	64
compensation_ok	59
be_willing	55
cabin_change	52
multi_change	52
rebook_pay	45
rank_pick	45
flight_change	39
be_nooffer	35
cabin_partial	26
upgrade_then_change	26
passenger_count_denied	25
refuse_then_book	18
flight_change_dest_first	16
two_denied	15
booking_by_rule	14
upgrade_and_bags	14
booking_round_trip	13
baggage_remove_denied	8
booking_one_stop	7
flight_change_denied	6
compensation_denied	6
insurance_add_denied	6

training files

file	examples
sft/all_passing_conversations.jsonl	2427
sft/round1_train_mix.jsonl	701
sft/round1b_train_mix.jsonl	1167
sft/round2_train_mix.jsonl	1278
sft/round3_train_mix.jsonl	1172

*_train_mix.jsonl is the exact mix each training round saw. all_passing_conversations.jsonl is every conversation that passed, before mixing.

how it was cooked 🧑‍🍳

1. re-skin: start from the τ²-bench airline database, rename everyone and everything, and drop every user an official task touches.
2. generate: scripted scenario families, each with gold actions computed from the database. Every scenario is replayed against the real airline tools by a validator before it is used.
3. self-play: the model plays the agent, gpt-6.1-sol plays the customer. Only conversations that score reward 1.0 are kept.
4. train: LoRA rank 32, three rounds, merged into the weights.

The whole pipeline was built and run end to end by an AI agent (Claude Opus 5.5).

results it produced

τ²-bench airline, 50 tasks × 4 trials: pass^1 0.850 with working notes, 0.865 without.

fine print

• Training data: the model's own conversations that scored reward 1.0, on (a) synthetic airline scenarios generated over a re-skinned copy of the tau2-bench airline database (renamed users, reservations and flights; users referenced by any official task excluded) and (b) the 30 official airline train-split tasks. No official test-split task, and no conversation on one, was used as training data. The round-3 scenario families and the working notes were designed after reviewing failure modes on official tasks, including test-split tasks; the scenarios are new look-alike cases, not copies of test tasks. 30 of the 50 airline tasks (the train split) were seen in training.
• The customer side of every conversation was generated by gpt-6.1-sol (low reasoning). The agent side is the model's own output.
• Everything here is synthetic. No real people, bookings or payment details.

code

• code/README.md
• code/merge.py
• code/mutation_test.py
• code/oracle_vc_bench.py
• code/train_lora.py
• code/validate_r2.py
• code/validate_r3.py
• code/validate_vc_bench.py
• code/vc_bench.py
• code/vc_bench_r2.py
• code/vc_bench_r3.py
• code/vc_bench_r3b.py
• code/vc_install.py
• code/vc_sft.py