Skip to content

cli-reference.md

Latest commit

 

History

History
381 lines (257 loc) · 6.65 KB

cli-reference.md

File metadata and controls

381 lines (257 loc) · 6.65 KB
Raw

CLI Reference

flattune

Main CLI entry point.

flattune [OPTIONS] COMMAND [ARGS]...

Options

Option Description
--version Show version
--help Show help

Commands Overview

Command Description
build Extract data and generate dataset with intelligent planning
train Train a model using the dataset
merge Merge LoRA adapter with base model
export Export model to target format (GGUF/MLX/HF)
benchmark Benchmark the model
report Generate training and benchmark report
run Run the full pipeline
teach Knowledge & Skill Compiler (teach framework)

build

Extract data and generate dataset with intelligent planning.

flattune build CONFIG_FILE [OPTIONS]

Options:

Option Description
-y, --yes Skip interactive approval, use suggested types
--source TEXT Source file/directory (overrides config)
-t, --types TEXT Dataset types to generate (can specify multiple)

Examples:

# Non-interactive (auto-accept suggested types)
flattune build configs/earthquake-qa.yml --yes

# Explicit dataset types
flattune build configs/earthquake-qa.yml --types facts --types context_qa

# Interactive mode
flattune build configs/earthquake-qa.yml

train

Train a model using the dataset.

flattune train CONFIG_FILE

Example:

flattune train configs/my_project.yml

merge

Merge LoRA adapter with base model.

flattune merge CONFIG_FILE

Example:

flattune merge configs/my_project.yml

export

Export model to target format (GGUF/MLX/HF).

flattune export CONFIG_FILE

Example:

flattune export configs/my_project.yml

benchmark

Benchmark the model using LM Studio, Ollama, or Transformers.

flattune benchmark CONFIG_FILE

Example:

flattune benchmark configs/my_project.yml

report

Generate training and benchmark report.

flattune report CONFIG_FILE

Example:

flattune report configs/my_project.yml

run

Run the full pipeline: extract → generate → train → merge → export → benchmark → report.

flattune run CONFIG_FILE

Example:

flattune run configs/my_project.yml

teach

The teach command group provides the Knowledge & Skill Compiler functionality.

flattune teach [OPTIONS] COMMAND [ARGS]...

teach knowledge

Teach from documents (Markdown, TXT, PDF, HTML, CSV, JSON, FlatSeek).

flattune teach knowledge SOURCES... [OPTIONS]

Options:

Option Description
--distill Use template-based generation (no LLM)
--teacher {openai,anthropic,ollama} LLM teacher for generation
--model TEXT Teacher model name (default: gpt-4o-mini)
--output, -o FILE Output file
--sample-types, -s TEXT Sample types to generate
--max-samples N Maximum samples to generate

Examples:

# Distill mode (no LLM required)
flattune teach knowledge ./docs/*.md --distill -o dataset.jsonl

# With OpenAI teacher
flattune teach knowledge ./docs/*.md --teacher openai -o dataset.jsonl

# With Ollama (local)
flattune teach knowledge ./docs/*.md --teacher ollama --model llama3.2 -o dataset.jsonl

teach openapi

Teach from OpenAPI/Swagger specifications.

flattune teach openapi SPEC [OPTIONS]

Options:

Option Description
--distill Use template-based generation
--teacher TEXT LLM teacher for generation
--output, -o FILE Output file

Example:

flattune teach openapi api.json --distill -o api_dataset.jsonl

teach database

Teach from database schemas.

flattune teach database SCHEMA [OPTIONS]

Options:

Option Description
--distill Use template-based generation
--teacher TEXT LLM teacher for generation
--output, -o FILE Output file

Example:

flattune teach database schema.sql --distill -o sql_dataset.jsonl

teach software

Teach from software APIs and tools.

flattune teach software SOURCES... [OPTIONS]

Options:

Option Description
--distill Use template-based generation
--teacher TEXT LLM teacher for generation
--sample-types, -s TEXT Sample types to generate

teach mcp

Teach from MCP server definitions.

flattune teach mcp SERVER [OPTIONS]

Options:

Option Description
--distill Use template-based generation
--teacher TEXT LLM teacher for generation
--output, -o FILE Output file

teach list-parsers

List available parsers.

flattune teach list-parsers

teach list-teachers

List available teacher models.

flattune teach list-teachers

teach list-generators

List available generators.

flattune teach list-generators

Makefile Commands

make help      # Show help
make install   # Install dependencies (pip install -e .)
make build     # Extract + generate dataset
make train     # Train model
make merge     # Merge LoRA
make export    # Export model
make benchmark # Run benchmarks
make report    # Generate report
make run       # Full pipeline
make test      # Run pytest
make lint     # Run ruff linter
make clean     # Clean outputs

Examples

Teach Framework (No Flatseek Required)

# Documents → dataset (no LLM)
flattune teach knowledge ./docs/*.md --distill -o dataset.jsonl

# OpenAPI → tool-calling dataset
flattune teach openapi api.json --distill -o api_dataset.jsonl

# SQL schema → NL-to-SQL dataset
flattune teach database schema.sql --distill -o sql_dataset.jsonl

# With LLM teacher
flattune teach knowledge ./docs/*.md --teacher openai -o dataset.jsonl

Build Pipeline (Intelligent Planning)

# Non-interactive with auto-detection
flattune build configs/earthquake-qa.yml --yes

# Explicit types
flattune build configs/earthquake-qa.yml --types facts --types context_qa

# Interactive planning
flattune build configs/earthquake-qa.yml

Full Traditional Pipeline

# Run full pipeline
flattune run configs/my_project.yml

# Or step by step
flattune build configs/my_project.yml
flattune train configs/my_project.yml
flattune merge configs/my_project.yml
flattune export configs/my_project.yml
flattune benchmark configs/my_project.yml
flattune report configs/my_project.yml