chirag070901/jira-agent-trajectories

Jira Agent Trajectories

Synthetic expert demonstrations generated by the Jira Env Simulation environment. The examples teach a language-model policy to select structured actions while triaging and resolving Jira-style ticket queues.

Dataset structure

Each row represents one decision in a multi-step episode:

• messages: system prompt, current Jira observation, and expert JSON action
• task_id: easy, medium, or hard
• seed: reproducible scenario seed
• step: decision number within the episode
• reward: post-action environment reward stored as metadata, not exposed in the model's input

The environment supports five actions: assign_ticket, start_work, resolve_ticket, change_priority, and add_comment. This expert-demonstration version contains the productive assign_ticket, start_work, and resolve_ticket paths. The lower-value administrative actions remain available to later reinforcement-learning rollouts and are scored by the environment.

Generation and splits

All examples are produced by a deterministic rule-based expert interacting with the same state machine used for evaluation. No real Jira tickets, customer data, or personal information are included.

• Training seeds: 0-999
• Validation seeds: 1000-1199
• Test seeds: 2000-2199

Split	Episodes	Decision examples
train	300	2,500
validation	60	500
test	60	500

The uploaded manifest.json records generation counts and statistics.json records schema validation, distributions, and cross-split duplicate checks.

Intended use

The dataset is intended as a supervised cold start for an open-weight Jira agent before reinforcement learning with environment rewards. It can also be used for structured-action generation and agent evaluation experiments.

Limitations

The queues are synthetic and simplified. The expert follows the environment's designed workflow and is not evidence of optimal behavior in every real Jira organization. Test examples should remain excluded from model training and hyperparameter selection.