Export tool call inputs and outputs

[woct0rdho]

All tool call inputs and outputs for Claude Code/Codex/Gemini CLI/OpenCode are exported, without truncation.

There can be entire file contents and terminal logs in the tool call outputs. All tool call inputs and outputs are anonymized in the same way as the chat inputs and outputs.

For concern about dataset size, I guess we can trust the deduplication and compression on HuggingFace's backend.

Some researches like https://blog.sweep.dev/posts/oss-next-edit suggest that showing entire file contents before and after the edit is helpful in training code edit models.

[woct0rdho]

@lhoestq I've tried to upload the dataset with the new schema (see the new README.md) to https://huggingface.co/datasets/woctordho/dataclaw , and the webpage shows

The dataset viewer is not available for this split.
Job manager crashed while running this job (missing heartbeats).
Error code:   JobManagerCrashedError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Could you help with this? Is it because the jsonl is too long?

[lhoestq]

Thanks for reporting ! It looks like an issue with the datasets library that crashes with this dataset, I'll try to have a fix asap

[wjessup]

@woct0rdho can you also reexport your data now that #7 is merged?

[woct0rdho]

@wjessup My dataset above was exported with #7 merged. Do you still see duplicate thinking blocks in it?

[wjessup]

@woct0rdho no this is fixed. thank you.