Fine-tuning JSONL format for OpenAI, Gemini and Mistral

{"messages":[{"role":"system","content":"Classify the support ticket. Reply with one label: billing, bug, feature_request or other."},{"role":"user","content":"I was charged twice for my March invoice."},{"role":"assistant","content":"billing"}]}
{"messages":[{"role":"system","content":"Marv is a factual chatbot that is also sarcastic."},{"role":"user","content":"What's the capital of France?"},{"role":"assistant","content":"Paris","weight":0},{"role":"user","content":"Can you be more sarcastic?"},{"role":"assistant","content":"Paris, as if everyone doesn't know that already.","weight":1}]}

{"messages":[{"role":"user","content":"What is the weather in San Francisco?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"San Francisco, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}

{"input":{"messages":[{"role":"user","content":"Hello, can you tell me how cold San Francisco is today?"}],"tools":[],"parallel_tool_calls":true},"preferred_output":[{"role":"assistant","content":"Today in San Francisco, it is not quite cold as expected. Morning clouds will give away to sunshine, with a high near 68°F (20°C) and a low around 57°F (14°C)."}],"non_preferred_output":[{"role":"assistant","content":"It is not particularly cold in San Francisco today."}]}

{"systemInstruction":{"role":"system","parts":[{"text":"Classify the support ticket. Reply with one label: billing, bug, feature_request or other."}]},"contents":[{"role":"user","parts":[{"text":"I was charged twice for my March invoice."}]},{"role":"model","parts":[{"text":"billing"}]}]}

# pip install tiktoken
import json
import sys

import tiktoken

ROLES = {"system", "user", "assistant", "tool"}  # Gemini files use "user" and "model" instead
MIN_EXAMPLES = 10        # OpenAI's documented minimum
MAX_TOKENS = 65_536      # OpenAI's per-example limit for the GPT-4.1 family
enc = tiktoken.get_encoding("o200k_base")

def problems(example):
    if not isinstance(example, dict):
        return ["line is not a JSON object"]
    messages = example.get("messages")
    if not isinstance(messages, list) or not messages:
        return ["missing a non-empty \"messages\" list"]
    found = []
    for i, m in enumerate(messages):
        role = m.get("role") if isinstance(m, dict) else None
        if role not in ROLES:
            found.append(f"message {i}: unknown role {role!r}")
        elif not isinstance(m.get("content"), (str, list)) and not m.get("tool_calls"):
            found.append(f"message {i}: content must be a string")
        if isinstance(m, dict) and m.get("weight", 1) not in (0, 1):
            found.append(f"message {i}: weight must be 0 or 1")
    if not any(isinstance(m, dict) and m.get("role") == "assistant" for m in messages):
        found.append("no assistant message to learn from")
    return found

path = sys.argv[1]
raw = open(path, "rb").read()
if raw.startswith(b"\xef\xbb\xbf"):
    print("file starts with a UTF-8 byte order mark: save as UTF-8 without BOM")
text = raw.decode("utf-8-sig")  # raises on bytes that aren't UTF-8

count, errors = 0, 0
for n, line in enumerate(text.split("\n"), start=1):
    if not line.strip():
        if n < text.count("\n") + 1:
            print(f"line {n}: blank line"); errors += 1
        continue
    try:
        example = json.loads(line)
    except json.JSONDecodeError as e:
        print(f"line {n}: invalid JSON ({e.msg} at column {e.colno})"); errors += 1
        continue
    count += 1
    for p in problems(example):
        print(f"line {n}: {p}"); errors += 1
    tokens = len(enc.encode(line))  # rough: counts the JSON syntax too
    if tokens > MAX_TOKENS:
        print(f"line {n}: about {tokens:,} tokens, over {MAX_TOKENS:,} (will be truncated)")

if count < MIN_EXAMPLES:
    print(f"only {count} examples: OpenAI needs at least {MIN_EXAMPLES}")
print(f"{count} examples, {errors} errors")

file starts with a UTF-8 byte order mark: save as UTF-8 without BOM
line 2: blank line
line 3: invalid JSON (Expecting value at column 97)
line 4: message 1: unknown role 'model'
line 4: no assistant message to learn from
line 5: invalid JSON (Expecting property name enclosed in double quotes at column 2)
line 6: invalid JSON (Extra data at column 13)
line 7: invalid JSON (Expecting value at column 1)
line 8: no assistant message to learn from
line 9: invalid JSON (Expecting property name enclosed in double quotes at column 2)
only 3 examples: OpenAI needs at least 10
3 examples, 9 errors

import csv
import json

SYSTEM = "Classify the support ticket. Reply with one label: billing, bug, feature_request or other."

# newline="" lets the csv module read line breaks inside quoted cells;
# utf-8-sig also accepts files that Excel saved with a byte order mark.
with open("tickets.csv", newline="", encoding="utf-8-sig") as src, \
     open("train.jsonl", "w", encoding="utf-8", newline="\n") as out:
    for row in csv.DictReader(src):
        ticket = row["ticket"].replace("\r\n", "\n").strip()
        example = {"messages": [
            {"role": "system", "content": SYSTEM},
            {"role": "user", "content": ticket},
            {"role": "assistant", "content": row["label"].strip()},
        ]}
        # one object per line; ensure_ascii=False writes "é" as UTF-8, not an escape
        out.write(json.dumps(example, ensure_ascii=False, separators=(",", ":")) + "\n")

// npm install csv-parse
import { readFileSync, writeFileSync } from "node:fs";
import { parse } from "csv-parse/sync";

const SYSTEM = "Classify the support ticket. Reply with one label: billing, bug, feature_request or other.";

// columns: true gives one object per row; bom: true drops an Excel byte order mark
const rows = parse(readFileSync("tickets.csv", "utf8"), { columns: true, bom: true, skip_empty_lines: true });

const lines = rows.map((row) =>
  JSON.stringify({
    messages: [
      { role: "system", content: SYSTEM },
      { role: "user", content: row.ticket.replace(/\r\n/g, "\n").trim() },
      { role: "assistant", content: row.label.trim() },
    ],
  }),
);
writeFileSync("train.jsonl", lines.join("\n") + "\n", "utf8");

{"messages":[{"role":"system","content":"Classify the support ticket. Reply with one label: billing, bug, feature_request or other."},{"role":"user","content":"Could you add dark mode?\nMy eyes would thank you."},{"role":"assistant","content":"feature_request"}]}