Omitting too much data?

[wendlerc]

Hi,

I just wanted to start a discussion here about the topic of which data is worth keeping / deleting.

It seems to me that the current approach omits data a little bit too generously. E.g., detailed tool outputs seem useful as well as subagent logs if someone wanted to use this data for (1.) mining of successful model rollouts (2.) subsequent tuning on those.

If one of the usecases is to make this data useful for model training the rollouts should be as complete as possible right? E.g., when tuning a model to leverage tool calls one could imagine to teacher force on all the tokens that lead up to a tool-call, then mask the output of the tool-call from the loss computation, and continue loss computation on the assistant response following the tool-call. For this one would want the entire tool-call in the dataset.

Best,
Chris.

[woct0rdho]

Detailed tool outputs should be exported in the latest version 0.3.2 . You can convert the json to yaml using the script in #1 for easier reading.

What coding agent do you use? Can you show an example of what is missing?

I haven't checked subagent logs. Can you also show an example of what is missing?

[woct0rdho]

Some missing data have been fixed in the main branch, mostly for tool outputs and subagents in Claude Code, and also image inputs in all coding agents. Could you try it again?