Intended Use
- Route machine questions to the smallest useful Machina MCP tool.
- Keep specialist model outputs separate from language-model reasoning.
- Produce concise engineering responses that separate observations from
hypotheses.
- Support local, sovereign, and edge-oriented deployments after quantization.
Training
- Base model:
mistralai/Mistral-7B-Instruct-v0.3
- Method: QLoRA, NF4 4-bit loading, bf16 compute
- Adapter: LoRA on attention and MLP projection modules
- Dataset: synthetic-but-grounded Machina MCP routing examples generated from
the open-source harness tool schemas
The training examples teach tool selection and Clertree response style. Runtime
facts must still come from telemetry, registered model plugins, maintenance
knowledge, and MCP tool results.
Quick Start
from machina_harness.agent import generate_agent_turn, load_agent
tokenizer, model = load_agent(adapter="clerktree/machina-agent-mistral-7b-lora")
response = generate_agent_turn("which model plugins are installed?", tokenizer, model)
print(response.text)
print(response.tool_calls)
Safety
Machina is decision support for engineers and operators. It is not a safety
controller. Validate outputs against site procedures, calibrated sensors, and
qualified human inspection before maintenance action.