teacher57/tau-retail-distillation

τ-bench retail distillation: teacher rollouts, SFT data and test conversations

Data behind the experiment "distill a Qwen3-32B teacher into Qwen3-14B on τ-bench retail". Model: teacher57/qwen3-14b-tau-distilled-from-32b. Code and write-up: GitHub, branch distillation-experiment.

Result in one line: the fine-tuned 14B scores +3.9 points pass^1 over a same-pod control on the retail test (45.7% vs 41.7%), which is not statistically significant (permutation p = 0.15).

Contents

path	what
teacher/teacher_rollouts_raw.jsonl	every teacher (Qwen3-32B-AWQ) attempt on the 114 hard train tasks: task_index, trial, reward, messages, seconds, ... (473 rows; rows for samples not run after a task already had 2 passes are marked skipped and have no conversation)
teacher/teacher32b_passed_sft.json	the 180 passing conversations as SFT data: {messages, tools}
teacher/teacher32b_passed_sft_reasoning.json	the same with the teacher's <think> reasoning kept
teacher/teacher32b_passed_meta.json	per-conversation stats (turns, tool calls, write calls, 14B group)
teacher/groups.json, hard_tasks.json, tools.json, retail_test_tasks.json	task groups (how the 14B did before), the 114 hard tasks, the tool schema, the 115 test tasks
retail_eval_conversations/retail_new_model_4trials_*	all 460 test conversations of the fine-tuned model (115 tasks × 4 trials): *_transcripts.jsonl = {task_index, trial, reward, messages}, *_results.jsonl = per-rollout reward and timing
retail_eval_conversations/retail_starting_adapter_rerun_4trials_*	the same for the starting adapter (control), run on the same pod with identical settings
results/distill_run.json, results/teacher_run.json	aggregated results (pass^k, per-trial, per-type, comparisons, training curves) and per-task teacher results

All evaluation settings: 115 retail test tasks, 4 trials, temperature 0.7, tool-calling agent, max 25 steps, GPT-4o as the simulated customer (the user turns in these conversations were generated by GPT-4o; check OpenAI's terms before using them to train other models), reward = final database state match.

Graphs

The teacher solved 86 of 114 hard tasks; panels split by how the 14B did before, passes per task, and passes over time.

Loss 0.53 → 0.35 over 384 steps; dotted lines are saved checkpoints.

Running pass^1, pass^1 to pass^4 per model, and pass^1 per trial.

Combo tasks (several actions in one request) barely move; the other tasks are where the new model is ahead.

Per-task pass rates, bootstrap of the +3.9 point difference (zero is inside it), and effect sizes with 95% intervals.

Difference by task type (post hoc split), tasks gained and lost, conversation lengths.

Notes and caveats

• The conversations of an earlier baseline test (37.8%, 2 trials, other pod) were not kept, only its rewards.
• A partial airline run (10 rollouts) is not included.
• The test tasks are τ-bench's public retail test split; none of them were used for training.

---

Addendum: the lookup-and-confirmation augmentation experiment

New files, all described in GitHub section 11. Model: teacher57/qwen3-14b-tau-lookup-confirm-augmented.

Graphs of the augmentation experiment

Training loss (511 steps, 1 epoch), pass^k of the three models on τ-bench retail (4 trials) and of baseline vs augmented on τ³ retail (default settings, 2 trials), and the causes of failed τ-bench rollouts. None of the differences is statistically significant.

path	what
augmentation/teacher32b_passed_sft_lookups_confirm.json	the 178 teacher conversations rebuilt so that every order is opened first and every change is confirmed (710 lookups and 309 confirmations inserted by rule)
augmentation/habit_dataset.json	habit-only dataset, 898 conversations; supervise lists the only assistant turns that carry loss (lookup and confirmation turns, 3,830 in total); 178 real, 341 synthetic lookups, 379 synthetic confirmations built from customers no test task uses. Not trained yet
augmentation/*_stats.json	statistics of both datasets
retail_eval_conversations/retail_aug_4trials_*	all 460 τ-bench retail test conversations of the augmented model (115 tasks × 4 trials) and the per-rollout results
tau3_eval/tau3std_base_results.json, tau3std_aug_results.json	τ³ retail default settings (temperature 0, GPT-4.1 customer, 114 tasks × 2 trials): baseline (starting adapter) and augmented model, all 228 conversations each, with rewards and costs
tau3_eval/first_run_tau3_old_results.json, first_run_tau3_new_results.json	the first τ³ run (temperature 0.7, 1 trial): starting adapter and distilled model
results/	training curves and run status (aug_sft_run.json, aug_train_log.jsonl), tau3_std_run.json, tau3_run.json, failure_causes_auto.json (rule-based cause of every failed τ-bench rollout of the three models), new_model_failure_causes.json and old_model_failure_causes.json (hand-read causes of the failed τ³ tasks)
renders/	augmented_sft_dialogs.html (original vs augmented conversations, inserted turns highlighted), habit_dataset.html (trained turns highlighted), tau3_retail_dialogs.html (all 114 tasks, old vs new, with causes of failure)

Headline numbers (not statistically significant): τ-bench 4 trials, pass^1 control 41.7%, distilled 45.7%, augmented 44.6%; τ³ default settings, baseline 43.4% vs augmented 47.8% (p = 0.33), pass^2 25.4% vs 35.1% (p = 0.063). The user turns in these conversations were generated by GPT-4o (τ-bench) or GPT-4.1 (τ³); check OpenAI's terms before using them to train other models.

Addendum 2: earlier SFT sets and all other evaluation conversations

path	what
earlier_sft/qwen3_14b_retail_train_rollout_sft.json, earlier_sft/qwen3_14b_retail_train_augmented.json	the SFT sets of the earlier experiments (the model's own rollouts and their augmented version)
grpo_task_pools/	task pools used for the GRPO runs (combo_pool*.json, 16 training and 64 test tasks)
earlier_eval_conversations/14b_*_retail_full_115.json, 32b_retail_full_115.json, 32b_airline_full_50.json	full evaluation runs of the 14B variants and the 32B teacher (all conversations)
earlier_eval_conversations/grpo_probe_transcripts/	conversations of the 20-task probes during the GRPO runs
earlier_eval_conversations/benchmark_transcript_*.jsonl	early benchmark runs of older SFT experiments
earlier_eval_conversations/airline_partial_new_*	the 10 airline conversations of the distilled model

All conversations of the distillation and augmentation experiments are in retail_eval_conversations/ (τ-bench, 4 trials each: starting adapter, distilled, augmented) and tau3_eval/ (τ³). Everything is also in the GitHub repo.