Finetunify

A really beautiful TUI app for local synthetic data generation <3 (we love finetunes).

Current version: v0.1.0

---

Why Finetunify

• Generate train/eval/test splits fast with local Ollama models or OpenRouter
• Tool-calling friendly, JSON-first, and retry-safe
• Clean TUI workflow with previews, logs, and live progress

---

Requirements

• Rust + Cargo (for building from source or first npm install build)
• Ollama running locally (ollama serve) for local generation
• OpenRouter API key (optional) for cloud model distillation
• At least one local model (ollama pull llama3) or an OpenRouter model

---

Installation

npm i -g finetunify

Run:

finetunify

---

Run from source

cargo run

---

Config & storage

• Config lives in: ~/.finetunify/finetunify.config.json
• Outputs: default to ./output/ (relative to the directory you launch finetunify from)
• Checkpoints: ./output/.finetunify_checkpoint.json

---

Controls

• Tab switch focus (menu/content)
• Up/Down move selection
• Enter select page / edit field / run action
• Space toggle booleans
• Right open selected menu item
• Left go back
• a add example
• d delete example
• : command palette
• g start generation
• s stop generation
• r refresh model list
• q quit

---

Prompt templating

The global prompt template supports:

• {data_spec}
• {example}
• {split}
• {index}
• {total}

System prompts are split‑specific and editable in the UI.

---

Tool calling + format

• Global tools json should be a JSON array of tool definitions (passed to Ollama /api/chat as tools).
• Global format can be json or a JSON schema object to constrain output.

---

Providers (Local + OpenRouter)

You can switch between local Ollama and OpenRouter in Global settings.

OpenRouter options:

• OpenRouter API key (stored in config)
• OpenRouter concurrency (parallel requests)
• OpenRouter retries (HTTP retry attempts)

Model lists are fetched dynamically based on the selected provider.

---

Examples + retries

• Global examples accepts one example per line. Each generated record is seeded with one example (rotating).
• Global json retries controls how many times to retry if the model returns malformed JSON.
  • The parser strips code fences, extracts the first JSON object/array, and removes trailing commas.
  • Live preview updates every 5 generated items on the Generate page.

---

Command palette

Open with : and run quick commands like:

• generate
• stop
• refresh
• provider ollama|openrouter
• set openrouter-key <key>
• page models|splits|global|examples|logs|generate
• model <name>
• examples add <text> / examples clear

---

Output format

Each output line is JSONL (when Wrap output is enabled) and includes:

• id, split, model, system, user
• response.content (raw)
• response.content_json (parsed JSON, if any)
• response.tool_calls (if any)

If Wrap output is disabled, the raw model response (or tool calls) is written directly.

---

Troubleshooting

• No models listed: run ollama list and ensure ollama serve is running.
• Generation stalls: the app waits for Ollama response; stop will take effect right after the current request completes.
• JSON invalid: increase Global json retries or tighten Global format schema.

---

License

MIT