Data Preparation

Source: src/renderer/src/components/DataPreparation.tsx

Overview

Tools for converting raw data into fine-tuning datasets. Three modes: CSV to JSONL conversion, file preview, and MCP-based generation.

CSV Preview

Upload a CSV file and preview its contents. Configurable row limit (1-1000). Displays column headers and data in a table.

API: POST /api/preparation/preview

CSV to JSONL Conversion

Map CSV columns to the standard fine-tuning format:

Target Field	Required	Description
instruction	Yes	The instruction or question
input	No	Additional context or input
output	Yes	The expected response

Select source columns from the CSV headers. The converter produces a JSONL file where each line is:

{"instruction": "...", "input": "...", "output": "..."}

Invalid rows (missing required fields) are skipped. The result shows the number of converted vs skipped rows.

API: POST /api/preparation/convert

MCP Dataset Generation

Generate fine-tuning datasets from MCP tool schemas. This connects to a configured MCP server, discovers its tools, and creates training examples that teach a model how to use those tools.

For each tool discovered, two training entries are generated:

1. Tool call example: instruction describes a use case, output is a JSON tool call with name and parameters.
2. Capability description: instruction asks what the tool does, output describes the tool and its parameter schema.

Requires:

• A configured MCP server (see MCP Integration)
• A loaded model (model ID is sent but not currently used for generation)
• An output path for the JSONL file

API: POST /api/preparation/generate-mcp

Backend

Implementation: backend/app/preparation/service.py

• preview_csv() — reads CSV with pandas, returns rows as dicts
• convert_csv_to_jsonl() — maps columns, validates rows, writes JSONL
• MCP generation is in backend/app/api/preparation.py (endpoint logic)