functiongemma
functiongemma-browser-edit-e14-4-scale-51200-datasetcode-tool-calling-functiongemma
Code Tool Calling Dataset (FunctionGemma Format)
A curated dataset of coding-focused tool-calling examples formatted for FunctionGemma-style training as demonstrated in google/mobile-actions.
Dataset Overview
Total: 82,406 examples (85% train / 15% validation split per subset)
Subsets
Config Name
Original Dataset
Original Subset
Train
Validation
Total
ToolRM
ibm-research/ToolRM-train-data
train
44,293
7,823
52,116
Toucan-1.5M-SFT… See the full description on the dataset page: https://huggingface.co/datasets/eacortes/code-tool-calling-functiongemma.function-gemma-tuning
FunctionGemma tool-calling SFT demo (Agents + Colab + HF Jobs + UV)
This repo is a FunctionGemma tool-calling fine-tune demo you can run:
On Colab: function_gemma_sft_tool_calling.ipynb
On HF Jobs: sft-tool-calling.py (standalone UV script for TRL SFT + LoRA/QLoRA)
Locally: sft-tool-calling.py (via uv run)
References:
Hugging Face Jobs guide
huggingface/skills
1) Train with a coding agent (HF Skills + Jobs)
Install Hugging Face Skills (Agent Context Protocol bundles)… See the full description on the dataset page: https://huggingface.co/datasets/burtenshaw/function-gemma-tuning.whispbook-functiongemma-speaker-attribution-sftfunctiongemma-finetune-dataset
FunctionGemma Finetune Dataset — Cardinal System Tools
Conversational traces for training lightweight models (FunctionGemma 270M) to translate natural language into executable function calls for system tools: clock, calendar, notes, translation, settings, and AI routing.
Dataset Format
JSONL, one sample per line. Format matches google/mobile-actions exactly.
Each sample has:
metadata: "train" or "eval" (90/10 split)
tools: A list of 7 available functions (random… See the full description on the dataset page: https://huggingface.co/datasets/SkGufranAhmed/functiongemma-finetune-dataset.functiongemma-thinking-curriculum
FunctionGemma Curriculum Dataset (Converted)
Converted from trace-based samples into FunctionGemma-ready chat format.
Schema
Each row contains:
messages: chat history with roles in {system,user,assistant,tool}
tools: function declarations for tokenizer.apply_chat_template(..., tools=...)
curriculum_stage: integer stage (1-5)
metadata fields: domain, subdomain, difficulty, format, quality_signals
Total rows: 1214
Stage counts
Stage 1: 280
Stage 2: 425… See the full description on the dataset page: https://huggingface.co/datasets/Cyleux/functiongemma-thinking-curriculum.