tau-bench
tunekv-taubench-data
TuneKV τ-bench GRPO data (§10)
Training and eval artifacts for the three τ-bench outcome-RL
(GRPO clip 0.2 + 0.1 reverse KL) arms. CE arms live in sibling subdirectories;
grpo_* subdirectories are the GRPO siblings (single variable vs CE = loss +
all-K-outcome data with per-task z-scored advantages).
Subdirectory
Arm
Rows
Key result (same stack)
grpo_airline30/
airline · Qwen3-Coder-30B
120 (K=4+K=8, 5/10 mixed)
0.46 → 0.42 full-50
grpo_retail30/
retail ·… See the full description on the dataset page: https://huggingface.co/datasets/Zhongzhu/tunekv-taubench-data.tau-bench-trajectories
AgentSuite/tau-bench-trajectories
Per-model agent trajectory data for tau-bench (public release).
Models: 30
Tasks per model: 165
One file per model: {model}.jsonl, one JSON object per line.
Fields: model_path, user_model_path, benchmark_name, task_name, sampling_params, user_sampling_params, messages, eval_result, meta.
sampling_params reflect each benchmark's own implementation; values the benchmark leaves unset are recorded as null (provider default).
Models… See the full description on the dataset page: https://huggingface.co/datasets/AgentSuite/tau-bench-trajectories.tau_bench
tau-bench trajectory and grader rows
This dataset contains Ergon-native sharded rollout-card exports for the paper artifact.
Evidence type: Full multi-step trace.
This export includes inspectable multi-step process evidence for rollout-card review.
Source URL: https://github.com/sierra-research/tau-bench/tree/59a200c6d575d595120f1cb70fea53cef0632f6b/historical_trajectories
License: See upstream tau-bench repository
Redistribution class: metadata-plus-fetch
Card claim class:… See the full description on the dataset page: https://huggingface.co/datasets/annon124816/tau_bench.apigen-tau-bench-split-turntaubench_traces_training_data
TauBench Traces Training Data
This dataset contains traces of conversations between a tool-using AI agent and users, formatted for fine-tuning.
Dataset Structure
The data is organized in JSONL format, where each line contains a conversation in the following structure:
{
"messages": [
{"role": "system", "content": "..."},
{"role": "user", "content": "..."},
{"role": "assistant", "content": "...", "tool_calls": [...]},
{"role": "tool", "tool_call_id":… See the full description on the dataset page: https://huggingface.co/datasets/jkazdan/taubench_traces_training_data.tau-bench-synthetic
tau-bench-synthetic
Synthetic tool-use training data for tau-bench, generated using a GT-first task construction pipeline with GLM-5 (via Fireworks API) as the trajectory generator.
Overview
This dataset was built to train small LLMs (e.g., Qwen3-1.7B) on multi-turn tool-use tasks without using the original tau-bench evaluation set. The pipeline follows a GT-first approach: ground-truth actions are constructed programmatically from the database, then an LLM generates… See the full description on the dataset page: https://huggingface.co/datasets/fuvty/tau-bench-synthetic.