Llama Fine-Tune JSONL Validator

updated 17 May 2026 · Llama 3.x / Llama 4 · ChatML · ShareGPT · Alpaca · separate pages for OpenAI, Anthropic, Gemini, Mistral

Llama fine-tune JSONL validator. Validates the three dataset shapes the open-source training stack actually uses: ChatML messages (Axolotl chat_template, Unsloth, TRL SFTTrainer, Together, Fireworks, Replicate, LLaMA-Factory, vLLM-LoRA), ShareGPT (legacy Axolotl, community datasets like OpenHermes / WizardLM / Glaive), and Alpaca (single-turn instruction tuning — Stanford Alpaca, Stanford Dolly, classic Q&A). Maps every error to the exact string Axolotl, Unsloth, or TRL print at startup.

Your training data never leaves this tab. The training platforms upload your file when a job starts; this pre-flight check is fully local.

⌨ Prefer the terminal? jsonlkit validate --llama data.jsonl · --sharegpt · --alpaca — same checks, in a pipe.

Which format does my trainer want?

The three formats, side by side

ChatML messages (the 2026 default)

{"messages":[
  {"role":"system","content":"You are a precise SQL generator."},
  {"role":"user","content":"Users signed up last month?"},
  {"role":"assistant","content":"SELECT * FROM users WHERE created_at >= date_trunc('month', now() - interval '1 month');"}
]}

ShareGPT (legacy)

{"conversations":[
  {"from":"system","value":"You are helpful."},
  {"from":"human","value":"Hi"},
  {"from":"gpt","value":"Hello"}
]}

Roles are renamed: human ↔ user, gpt ↔ assistant, system stays. Content is in value, not content.

Alpaca (single-turn instruction tuning)

{"instruction":"Translate to French.","input":"Good morning","output":"Bonjour"}

input is optional — many Alpaca examples leave it empty when the prompt is self-contained. Each row is one independent training example; Alpaca doesn't support multi-turn conversations.

Tool calls (ChatML)

{"messages":[
  {"role":"user","content":"Weather in Paris?"},
  {"role":"assistant","content":null,"tool_calls":[
    {"id":"c1","type":"function","function":{"name":"get_weather","arguments":"{\"city\":\"Paris\"}"}}
  ]},
  {"role":"tool","tool_call_id":"c1","content":"18C, clear"},
  {"role":"assistant","content":"It's 18 C and clear in Paris."}
]}

DPO / preference (chosen / rejected)

{"prompt":"Summarize: ...","chosen":"Better answer.","rejected":"Worse answer."}

Used by TRL's DPOTrainer and Together's preference fine-tuning. Each row is one prompt with a preferred and rejected completion. Not validated as a primary mode on this page — use the ChatML mode and treat prompt as the user turn.

Which Llama model? It doesn't matter for the JSONL

The chat template that wraps role/content into <|begin_of_text|>, <|start_header_id|>, etc. is applied by the trainer at tokenization time. Don't put those tokens in your JSONL. Just write clean role/content pairs and let the trainer's chat template do the wrapping.

What this validator checks

• ChatML mode: messages array with valid roles (system, user, assistant, tool), at least one assistant turn, non-empty content (or valid tool_calls with content: null), no mixing with conversations.
• ShareGPT mode: conversations array with valid from values (system, human, gpt, tool), at least one gpt turn, non-empty value, no mixing with messages.
• Alpaca mode: non-empty instruction string, optional input string, non-empty output string. No extra top-level keys allowed.
• Each line is valid JSON, one object per line.
• UTF-8 encoding, no BOM, no trailing-only newline.

Common mistakes this validator catches

{"conversations":[{"from":"user","value":"Hi"},{"from":"assistant","value":"Hi"}]}
// Error: ShareGPT uses 'human'/'gpt', not 'user'/'assistant'.

{"messages":[{"role":"human","content":"Hi"}]}
// Error: ChatML uses 'user', not 'human'.

{"conversations":[{"from":"human","content":"Hi"}]}
// Error: ShareGPT uses 'value', not 'content'.

{"instruction":"Do X.","output":""}
// Error: Alpaca 'output' must be non-empty — it's the training target.

{"messages":[...], "conversations":[...]}
// Error: don't mix ChatML and ShareGPT in the same record.

Real error strings from the OSS training stack

How to use it

1. Pick the format your trainer expects. ChatML is the right default for 2026.
2. Drop a .jsonl file or paste records.
3. Click Validate. Each line is checked against the rules for the selected format.
4. Inspect the Error List — line number and reason.
5. Download valid examples only rebuilds a clean file with broken lines stripped.
6. Need to switch formats? The fine-tune editor converts ChatML ↔ ShareGPT in one click.

Tips & common pitfalls

• Don't put chat-template tokens in your JSONL. The <|begin_of_text|>, <|start_header_id|>user<|end_header_id|>, <|eot_id|> markers are added by the trainer's tokenizer. If you bake them in yourself, you'll double-wrap and the model will train on the literal tokens.
• ShareGPT in Axolotl 0.6+ is deprecated. The format still parses but only via the new chat_template path. New projects should use ChatML.
• Alpaca is single-turn only. If your dataset has multi-turn conversations, use ChatML — Alpaca can only express instruction → output.
• Llama version doesn't change the JSONL. 3.1, 3.2, 3.3, 4 — all the same shape. The trainer picks the right chat template based on the model.
• Don't mix formats in one file. Trainers iterate line by line and assume a uniform shape. Mixed files fail with cryptic indexing errors.
• Empty strings train on whitespace. The trainer may not error, but you'll waste compute and the model will pick up the noise. This validator flags empties.

Troubleshooting

Frequently asked questions

Related tools

• OpenAI Fine-Tune ValidatorChat messages + tool_calls, vision, developer role, weight.
• Anthropic Fine-Tune ValidatorBedrock Claude 3 Haiku — top-level system.
• Gemini Fine-Tune ValidatorVertex AI contents + parts, function calling.
• Mistral Fine-Tune ValidatorOnline alternative to validate_data.py.
• Fine-tune dataset editorConvert ChatML ↔ ShareGPT ↔ Alpaca in one click.
• Train / val / test splitterRandom 80/10/10 split of a clean JSONL.
• Token counterPer-line token estimate for cost planning.