data
Tool Calling SFT Mix
Supervised fine-tuning dataset for tool use, unifying several public tool-calling corpora plus deliberate no-call examples under one schema.
About Tool Calling SFT Mix
A published Hugging Face dataset that merges ToolBench, xLAM60k, OpenFunctions v1 and others into a single schema (tools_json, messages_json, target_json, meta_source, n_calls), then mixes in non-tool instruction and plain-text examples from Dolly and WikiText to guard against catastrophic forgetting of general language ability during tool-use fine-tuning.
Read
- write-up
Falcon Twig: a tool-calling fine-tune that lost to the base model
I fine-tuned Falcon-H1-7B-Instruct on a tool-calling mix with QLoRA and re-ran BFCL. It scored 24.50% live accuracy against the base model's 67.21%.