Prepare CSV Data for LLM Fine-Tuning and Eval Sets

CSV as a Fine-Tuning Source of Truth

Many teams author examples in spreadsheets, export CSV, then convert to JSONL for trainers. Keeping a clean CSV (or converting back to CSV for QA) makes review, deduplication, and eval scoring much easier than editing giant JSONL files by hand.

Choose a Row Schema

Pick one pattern and stick to it across train and eval.

id,split,prompt,completion
1,train,"Classify sentiment: Great battery life","positive"
2,train,"Classify sentiment: Broke in a week","negative"
3,valid,"Classify sentiment: It's fine","neutral"

Step-by-Step: Build the Dataset

Quality beats quantity—especially early.

• Collect examples in Excel or Google Sheets with frozen headers.
• Export UTF-8 CSV (Excel to CSV).
• Clean in the Online CSV Editor: quotes, blank rows, label enums.
• Add id and split columns.
• Convert to the trainer’s JSON/JSONL shape with CSV to JSON or a script.
• Keep the CSV as the editable source; regenerate JSONL as needed.

From CSV to JSONL

Trainers often want one JSON object per line.

import csv, json

with open("train.csv", newline="", encoding="utf-8") as f, open("train.jsonl", "w", encoding="utf-8") as out:
    for row in csv.DictReader(f):
        if row.get("split") != "train":
            continue
        obj = {
            "messages": [
                {"role": "user", "content": row["prompt"]},
                {"role": "assistant", "content": row["completion"]},
            ]
        }
        out.write(json.dumps(obj, ensure_ascii=False) + "\n")

Eval Sets Deserve the Same Care

Fine-tuning without a frozen eval set is just vibes. Hold out examples that represent production traffic, including hard cases.

• Never train on eval rows.
• Store expected outputs for exact-match or rubric scoring.
• Version eval CSV alongside model versions.
• Include adversarial and empty-input cases.

Quality Bar for Each Row

Reject examples that teach the wrong behavior.

• Instructions in the prompt match the completion style.
• No contradictory labels for near-duplicate inputs.
• No PII unless the environment is approved.
• Completions are the behavior you want in production—not jokes or hedges.
• Quoting in CSV does not corrupt multiline prompts.

Real-World Examples

How teams actually manage fine-tune CSVs.

Common Mistakes

Dataset bugs dominate "fine-tune failed" stories.

• Train/eval contamination.
• Inconsistent label strings.
• Multiline fields broken by bad CSV export.
• Too few examples for the diversity of production inputs.
• Editing JSONL by hand until nobody trusts the source.

Best Practices

Treat datasets like code.

• CSV (or a database) as source of truth; JSONL as build artifact.
• ids + splits on every row.
• Lint for blank completions and illegal labels in CI.
• Review a random sample every release.
• Document the intended use and refusal behavior.

Why Use Convert CSV Online?

Convert CSV Online is free, browser-based, and requires no account for everyday conversions. Clean fine-tune tables, move between CSV and JSON for trainers, and spot quoting issues before a paid training run. Client-side workflows work on Windows, macOS, and Linux browsers.

Conclusion

Fine-tuning success is mostly dataset discipline: clear columns, clean splits, JSONL as an artifact, and eval sets you refuse to contaminate. CSV is a solid human-friendly source of truth.