Hi, thanks for releasing A3-Synth.
I’m trying to reproduce the data generation pipeline and got a bit stuck at the step between exploration and task intent generation.
I found:
- the exploration prompt in
agent_as_annotators/prompts/__init__.py
TASK_INTENT_PROMPT_TEMPLATE
scripts/generate_task_intents.py, which consumes prepared prompt files from outputs/task_intents/prompts/...
But I couldn’t find the script that creates those prompt files from the exploration trajectories.
Could you clarify how this intermediate step was done?
Specifically:
- What part of the exploration trajectory is passed to the Task Generator — full trajectory, selected step, trajectory prefix, chat messages, AXTree/screenshot, or something else?
- How are the
exploration_step_num values selected?
I’m asking because this step seems important for understanding how the exploration phase is turned into grounded task intents.
Thanks!
Hi, thanks for releasing A3-Synth.
I’m trying to reproduce the data generation pipeline and got a bit stuck at the step between exploration and task intent generation.
I found:
agent_as_annotators/prompts/__init__.pyTASK_INTENT_PROMPT_TEMPLATEscripts/generate_task_intents.py, which consumes prepared prompt files fromoutputs/task_intents/prompts/...But I couldn’t find the script that creates those prompt files from the exploration trajectories.
Could you clarify how this intermediate step was done?
Specifically:
exploration_step_numvalues selected?I’m asking because this step seems important for understanding how the exploration phase is turned into grounded task intents.
Thanks!