Flattune

Compile knowledge into language models.

Flattune transforms documentation, APIs, databases, and Flatseek indexes into training datasets, fine-tuned models, and benchmark reports.

---

The problems

Most frameworks expect you to already have a training dataset. But that's rarely what you actually have.

• You have documentation.
• You have API specifications.
• You have database schemas.
• You have internal wikis and years of accumulated knowledge.

Turning all of that into high-quality training data is usually the hardest part.

Knowledge
    │
    ▼
Export → Clean → Split → Reformat → Generate →
Train → Merge → Export → Benchmark

Every stage uses a different tool. Every project rebuilds the same pipeline. Flattune replaces that entire workflow with one configuration and one reproducible pipeline.

---

A knowledge compiler

Flattune doesn't start with datasets. It starts with knowledge.

Think of Flattune as a compiler. Instead of compiling source code into binaries, it compiles knowledge into language models.

The dataset isn't your starting point—it's an artifact generated from your knowledge.

---

Build once, run anywhere

Every stage of the pipeline is independent. Run everything with one command:

flattune run project.yml

Or execute individual stages:

flattune build project.yml      # Knowledge → Dataset
flattune train project.yml      # Dataset → LoRA
flattune merge project.yml      # LoRA → Base Model
flattune export project.yml     # GGUF / MLX / HuggingFace
flattune benchmark project.yml  # Performance & Quality

Each stage produces reusable artifacts that can be resumed independently.

---

Why Flatseek?

Flattune works with raw files, but Flatseek unlocks a different workflow. Without Flatseek, every model starts by scanning the entire knowledge base.

docs/
api/
wiki/
faq/
database/

        │

        ▼

Generate Dataset

        │

        ▼

Train Model

Want another model for a different domain?

Start over.

Scan everything again.

Generate another dataset.

Repeat.

---

Flatseek indexes your knowledge once.

After that, every model starts with a search query instead of a full scan.

Knowledge
      │
      ▼
Build Flatseek Index (once)
      │
      ▼
knowledge.fsk
      │
      ├── tags:python
      ├── tags:docker
      ├── tags:postgres
      ├── tags:kubernetes
      └── ...

Flattune simply asks Flatseek for the knowledge it needs.

The knowledge never moves.

Only the search query changes.

Build one index.

Train as many specialized models as you need.

---

Teach

Instead of manually writing instruction datasets, Teach analyzes your knowledge and generates dataset types that match the source.

Source	Generated Datasets
Markdown	Facts, Context QA, Procedures, Conversations
OpenAPI	Tool Calling, API QA, Conversations
SQL Schema	Text-to-SQL, Schema QA
FAQ	QA, Conversations, Summaries
Product Catalog	Attribute Extraction, Recommendations
Flatseek Index	Coding QA, Bug Fixes, Conversations, Code Completion

Teach supports two generation modes.

Distill Mode

No LLM required.

Knowledge
    │
    ▼
Rules + Templates
    │
    ▼
Training Dataset

Fast, deterministic, reproducible, and works completely offline.

flattune teach knowledge ./docs --distill

Perfect for documentation, APIs, SQL schemas, FAQs, and structured knowledge.

Teacher Mode

Use an LLM to enrich the generated datasets.

Knowledge
      │
      ▼
LLM Teacher
      │
      ▼
Synthetic Examples
Multi-turn Dialogue
Reasoning
Tool Calling
      │
      ▼
Training Dataset

flattune teach knowledge ./docs \
    --teacher openai

flattune teach knowledge ./docs \
    --teacher ollama \
    --model llama3.2

Teacher Mode doesn't replace your knowledge.

It expands it—creating additional instructions, conversations, and examples that improve coverage before fine-tuning.

---

One source, many datasets

Markdown

docs.md → Facts        {"instruction": "...", "response": "..."}
        → Context QA   {"question": "...", "answer": "..."}
        → Procedure    {"steps": [...]}

OpenAPI

api.json → Tool calling  {"tool": "create_invoice", "args": {...}}
         → API QA        {"question": "...", "answer": "..."}

SQL Schema

schema.sql → Text-to-SQL  {"question": "...", "sql": "SELECT ..."}
           → Schema QA    {"question": "...", "answer": "..."}

FAQ

faq.json → Conversational  {"question": "...", "answer": "..."}
         → Direct QA       {"question": "...", "answer": "..."}

Product Catalog

products.csv → Classification  {"input": "...", "label": "..."}
             → Recommendation  {"query": "...", "products": [...]}

Flatseek Index

knowledge.fsk → Coding QA           {"question": "...", "answer": "..."}
              → Bug Fix             {"error": "...", "fix": "..."}
              → Code Completion     {"prefix": "...", "completion": "..."}
              → Conversation        {"messages": [...]}
              → Instruction Following {"instruction": "...", "response": "..."}

---

Pipeline

flattune run configs/my_project.yml

or stage by stage:

flattune build configs/my_project.yml     # detect → plan → generate
flattune train configs/my_project.yml     # fine-tune
flattune merge configs/my_project.yml     # merge LoRA
flattune export configs/my_project.yml    # export
flattune benchmark configs/my_project.yml # benchmark

---

Intelligent Planner

Flattune analyzes your sources before generating anything.

$ flattune build docs/

Detected
✓ Markdown
✓ OpenAPI
✓ SQL

Recommended
✓ Facts
✓ Context QA
✓ Tool Calling
✓ NL→SQL

Estimated samples: 48,200

Continue? (Y/n)

No surprise dataset sizes. No wasted generation runs.

---

Features

Dataset types	17 built-in (facts, QA, tool calls, conversations, RAG, ...)
Source detection	Markdown, JSON, OpenAPI, SQL, CSV, HTML, Flatseek
Training backends	Unsloth (GPU), Transformers (CPU/MPS)
Export formats	GGUF, MLX, HuggingFace, Safetensors
Import targets	LM Studio, Ollama
Reproducibility	seeds, hashes, config snapshots, git info
Benchmarking	latency, throughput, quality, hallucination, regression

---

Installation

pip install -e .
pip install -e ".[unsloth]"   # GPU support

Python 3.10+ · 8GB+ VRAM for 4B models · 20GB+ disk

No GPU? Transformers backend runs on CPU and Apple MPS.

---

Quick start

Option 1: Use HuggingFace datasets (recommended)

Flattune can directly use public datasets from HuggingFace as .fsk files:

# Build dataset from HuggingFace .fsk (Olympic athletes example)
flattune build configs/athletes-qa.yml --yes

# Train the model
flattune train configs/athletes-qa.yml

# Merge LoRA with base model
flattune merge configs/athletes-qa.yml

# Export to GGUF
flattune export configs/athletes-qa.yml

# Run full pipeline
flattune run configs/athletes-qa.yml

Option 2: Teach from local documents

flattune teach knowledge ./docs/*.md --distill -o dataset.jsonl

Example config: HuggingFace .fsk dataset

# configs/athletes-qa.yml
name: athletes-qa

flatseek:
  path: https://huggingface.co/datasets/flatseek/public-dataset/resolve/main/271k-athletes.fsk
  query: "*"

dataset:
  type: qa
  query: "country:(China OR Japan OR Korea OR India)"
  max_samples: 500
  generators: [qa, facts]

model:
  source: huggingface
  repo: HuggingFaceTB/SmolLM2-135M

train:
  backend: transformers
  epochs: 1
  lora_rank: 8

benchmark:
  backend: lmstudio
  prompt_file: configs/athletes-prompts.json

Available public datasets on HuggingFace

• flatseek/public-dataset/271k-athletes.fsk - 271K Olympic athletes (1800-2000)
• More datasets coming soon

See Configuration Reference for all options.

---

Ecosystem

Flatseek	Search & index engine — stores knowledge
Flatvec	Semantic search — Vector Retreival, RAG
Flattune	Fine-tuning orchestrator — compiles knowledge into models

Each works standalone. Together, Flatseek feeds Flattune.

---

Documentation

• Getting Started
• Configuration Reference
• Dataset Types
• Teach Framework
• CLI Reference
• API Reference
• Benchmarking

---

Roadmap

In progress	Coming next
Source detectors (PDF, CSV)	Axolotl backend
Streaming generation	MLX-LM backend
flattune build --plan preview	LlamaFactory backend
Regression testing	Distributed / multi-GPU training
Semantic deduplication	Hyperparameter optimization

See TODO.md.

---

License

Apache License 2.0