How zhongyi-zhou/toolgrad works
ToolGrad is a tool-use workflow and dataset-generation framework, with separate tooling for preparing training data, fine-tuning models, and evaluating tool-use results. Its runtime workflow samples APIs, proposes and executes calls, selects reports, and updates a workflow; prebuilt graphs support MCP and ToolBench. The README documents raw workflow generation, SFT, and BFCL evaluation. Some large modules and data branches are indexed but not excerpted, so their internal call relationships are omitted rather than inferred.
Main entry points: graph_lib.py Workflow stages · module_lib.py Proposal and execution · toolbench_langchain_utils.py ToolBench adapter · toolgrad_on_mcp.py MCP graph
Components 23
Workflow runtime
- MCP graph
toolgrad/prebuilt/toolgrad_on_mcp.py - ToolBench graph
toolgrad/prebuilt/toolgrad_on_toolbench.py - Workflow stages
toolgrad/modules/graph_lib.py - Proposal and execution
toolgrad/modules/module_lib.py - Prompt construction
toolgrad/modules/prompt_lib.py - Workflow state
toolgrad/states.py - Run tracing
toolgrad/utils/trace_utils.py
Tool integrations
- LLM providers
toolgrad/utils/langchain.py - MCP tools
toolgrad/utils/mcp.py - ToolBench adapter
toolgrad/utils/toolbench/toolbench_langchain_utils.py - ToolBench catalog
toolgrad/utils/toolbench/toolbench_data_utils.py
Dataset preparation
- Workflow generation
src/generate_toolgrad_data.py - JSON data helpers
toolgrad/utils/data.py - Data formatting
src/data/data_format_lib.py - Negative sampling
src/data/negative_sample_lib.py - Dataset splitting
src/data/split_lib.py
Training and evaluation
- Model support
src/models/model_lib.py - SFT training
src/train/train_sft.py - Tool-use evaluation
src/eval.py - Evaluation metrics
src/eval_lib.py
Other components
- User query
- ToolBench API service
- MCP service
Connections 23
- User query → MCP graph: submits query
README.md - MCP graph → Workflow stages: configures stages
toolgrad/prebuilt/toolgrad_on_mcp.py - ToolBench graph → Workflow stages: configures stages
toolgrad/prebuilt/toolgrad_on_toolbench.py - Workflow stages → Proposal and execution: invokes
toolgrad/modules/graph_lib.py - Workflow stages → Workflow state: updates state
toolgrad/modules/graph_lib.py - Proposal and execution → Prompt construction: uses prompts
toolgrad/modules/module_lib.py - Proposal and execution → Workflow state: uses schemas
toolgrad/modules/module_lib.py - Proposal and execution → LLM providers: uses clients
toolgrad/modules/module_lib.py - Workflow stages → MCP tools: samples MCP APIs
toolgrad/modules/graph_lib.py - Workflow stages → Run tracing: records workflow
toolgrad/modules/graph_lib.py - MCP graph → MCP tools: configures service
toolgrad/prebuilt/toolgrad_on_mcp.py - MCP tools → MCP service: discovers tools
toolgrad/utils/mcp.py - ToolBench graph → ToolBench adapter: selects integration
- ToolBench adapter → ToolBench catalog: loads API metadata
toolgrad/utils/toolbench/toolbench_langchain_utils.py - ToolBench adapter → ToolBench API service: wraps API calls
- Workflow generation → LLM providers: uses model clients
src/generate_toolgrad_data.py - Workflow generation → Run tracing: records traces
src/generate_toolgrad_data.py - Data formatting → ToolBench adapter: reads tool names
src/data/data_format_lib.py - Data formatting → JSON data helpers: writes metadata
src/data/data_format_lib.py - Negative sampling → Data formatting: uses formatted data
src/data/negative_sample_lib.py - SFT training → Model support: trains model
- Tool-use evaluation → Evaluation metrics: creates judge
src/eval.py - Evaluation metrics → Model support: loads model support
src/eval_lib.py