SOB-FT-ToolACE

Mapped standalone corpus from Team-ACE/ToolACE.

See also the full mapping guide in README_toolace.md in the project repo.

What changed from the source

	Source	This repo
Format	system + conversations[]	SOB FT schema (null columns removed)
Tool calls	Bracket syntax [Func(a=1)]	Hermes <tool_call> blocks or JSON array (--ground-truth-format)
Tool schemas	Embedded in system prompt	json_schema OpenAI tools[] for invoked functions
Parameter types	Python-style (dict, int, float)	Draft 7 (object, integer, number) — 0 invalid after normalize
Conversation	Multi-turn chat	Prior turns → context; last user → question
Rows	11,300	11,300 (1:1)
Provenance	—	source_dataset=Team-ACE/ToolACE, record_id=toolace-{idx}

Key transformations

1. Role rename: user→human, assistant→gpt, keep tool
2. Training pair: last tool-call assistant turn + preceding user turn
3. Ground truth: parsed bracket calls → structured JSON/Hermes
4. Schema repair: normalize corrupted type fields; drop invalid types

All-null reserved columns are dropped.

Regenerate

python3 merge_toolace.py --dry-run
python3 merge_toolace.py
python3 merge_toolace.py --ground-truth-format json

Merge into extract

python3 merge_ft_corpus.py --corpus mariem123kfg/sob-ft-toolace --source-prefix Team-ACE/ToolACE