datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Tau2-Bench-Airline-With-Code-Agents
Dataset Card for a Code Agent Version of Tau Bench 2 Airline
Dataset Summary
This dataset includes sample traces and associated metadata from multi-turn interactions between an code agent and AI assistant. The dataset is based on the Airline environment from Tau^2 Bench and contains traces from both the original version and a version made at Snorkel AI using code agents to solve the same tasks (indicator in the version field; details below).
Curated by: Snorkel AI… See the full description on the dataset page: https://huggingface.co/datasets/snorkelai/Tau2-Bench-Airline-With-Code-Agents.airline-voice-concise
airline-voice-concise
Made with the whileai SDK · Used by: recipes/community/airline-voice-concise-under-probe-outcome-filter · Collection: Register
Training data for putting a speaking register into a model's weights. An
airline support agent that leads with the answer and stops, trained so the
register survives with no instruction in the prompt.
Trained on this set, Qwen3-4B goes from 2.2% to 92.1% of held-out replies
in the register, and becomes less likely to omit required… See the full description on the dataset page: https://huggingface.co/datasets/while-ai/airline-voice-concise.airline-resist-jailbreaks
airline-resist-jailbreaks
Made with the whileai SDK · Collection: Robustness
Jailbreak resistance for a customer support agent, trained on simulated
attacks and tested on real ones.
The real attacks come from elder-plinius/L1B3RT4S,
a public library of working jailbreaks. We read it to extract the attack
techniques and never trained on a single string from it. It is the
evaluation set, unseen by the model.
On 165 unseen blocks from a public jailbreak library the agent holds its… See the full description on the dataset page: https://huggingface.co/datasets/while-ai/airline-resist-jailbreaks.tau2-airline-deepseek-distill
τ²-bench airline · DeepSeek teacher trajectories
Successful multi-turn agent trajectories on τ²-bench's
airline domain, generated by running DeepSeek V4 Flash as the agent through the real τ²-bench
harness — same system prompt, same 14 tool schemas, same dialogue loop, same evaluator.
Used to behavior-clone the RL warm start
yuyu0529nya/qwen2.5-7b-tau2-airline-sft-lora,
which is the step 0 of the tau2_airline verl recipe.
Why these exist
GRPO on τ²-bench-airline… See the full description on the dataset page: https://huggingface.co/datasets/yuyu0529nya/tau2-airline-deepseek-distill.aprm-sft-thoughts-tau2-airline-policy_best-adamw30-lp0
Act-PRM SFT thoughts — tau2-bench airline
Act-PRM (Action Process Reward Models) infers the latent thoughts behind
logged, action-only agent demonstrations via an offline EM. For each
logged action x in state s we sample G=4 candidate thoughts z,
score each by the length-penalized action likelihood
reward(z) = p(x | s, z)
(len_frac grows with the thought's token length), and mark the best thought
(argmax reward). The (thought + action) span is then what downstream SFT / RL… See the full description on the dataset page: https://huggingface.co/datasets/mzio/aprm-sft-thoughts-tau2-airline-policy_best-adamw30-lp0.aprm-sft-thoughts-tau2-airline-base_best-adamw30-lp0
Act-PRM SFT thoughts — tau2-bench airline
Act-PRM (Action Process Reward Models) infers the latent thoughts behind
logged, action-only agent demonstrations via an offline EM. For each
logged action x in state s we sample G=4 candidate thoughts z,
score each by the length-penalized action likelihood
reward(z) = p(x | s, z)
(len_frac grows with the thought's token length), and mark the best thought
(argmax reward). The (thought + action) span is then what downstream SFT / RL… See the full description on the dataset page: https://huggingface.co/datasets/mzio/aprm-sft-thoughts-tau2-airline-base_best-adamw30-lp0.aprm-thought-generations-tau2-airline
Act-PRM Thought Generations — tau2-bench Airline
Latent thoughts inferred behind logged, action-only agent demonstrations by the
Act-PRM offline EM (Action Process Reward Models), for the tau2-bench airline
domain. For each logged (state s, action x) the EM-trained generator samples
G=4 candidate thoughts z, and each candidate is scored by the
length-penalized action likelihood of the logged action:
likelihood = p(x | s, z) # policy per-action-token… See the full description on the dataset page: https://huggingface.co/datasets/mzio/aprm-thought-generations-tau2-airline.tau-dev-task-airline-v1
tau-dev-task-airline-v1
Multi-turn tool-calling SFT dataset (268 records, 200 / 18 / 50 train / validation / test) derived from sierra-research/tau-bench airline-domain trajectories.
Meant to be used as a benchmark dataset for developing and validating data processing, training, and eval workflows involving tool use. Note: tau-bench is a widely-used public benchmark and many recently-trained models may have encountered variants of these trajectories during training, so be mindful of… See the full description on the dataset page: https://huggingface.co/datasets/lefft/tau-dev-task-airline-v1.