Mistral Fine-Tune JSONL Validator — Online Alternative to validate_data.py

updated 17 May 2026 · Mistral AI Studio · La Plateforme · Azure AI Foundry · mistral-finetune · separate pages for OpenAI, Anthropic, Gemini, Llama

Mistral fine-tune JSONL validator. Same checks as Mistral's official mistralai/mistral-finetune/utils/validate_data.py — no pip install mistral-finetune, no YAML, no upload. Validates Mistral AI Studio, La Plateforme, and Azure AI Foundry SFT datasets in four shapes: default Instruct (messages), pretrain (text), function-calling Instruct with the 9-character id/tool_call_id rule, and Pixtral image content. Catches UnrecognizedRoleError, InvalidAssistantMessageException, ToolCallFormatError, and the minimum-conversation/byte checks before La Plateforme rejects your file.

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

⌨ Prefer the terminal? jsonlkit validate --mistral data.jsonl — same checks, in a pipe.

Where can I fine-tune Mistral in 2026?

Four paths, same JSONL format under the hood:

• Mistral AI Studio — Mistral's hosted fine-tuning, successor to the La Plateforme fine-tune endpoint (now flagged deprecated in the public docs). Supports Mistral Small (25.x), Mistral Large 2411, Mistral Nemo, Ministral 3B / 8B, Codestral (25.08), Pixtral (12B / Large), Magistral. $4 minimum job fee, $2/month per stored model.
• Azure AI Foundry — Microsoft's managed Mistral fine-tuning. Supports Mistral Large 2411, Mistral Nemo, Ministral 3B.
• mistral-finetune (self-hosted LoRA) — the open-source repo at github.com/mistralai/mistral-finetune. Works with any v3-tokenizer-compatible weight: Mistral 7B v0.3, Mixtral 8x7B (after vocab extension), Mistral Nemo, Mistral Large v2, Codestral 22B.
• Third-party (community weights) — Together, Fireworks, RunPod, Unsloth, Axolotl all accept the same Mistral messages shape. For ShareGPT-style datasets see the Llama validator.

The four Mistral fine-tune formats

1. Default Instruct (messages)

{"messages":[
  {"role":"system","content":"You are a terse code reviewer."},
  {"role":"user","content":"Review: print('hi')"},
  {"role":"assistant","content":"Looks fine.","weight":1}
]}

Roles: system, user, assistant, tool. No developer role. weight: 0 on an assistant turn skips it from the training loss (useful when you want context without teaching imitation); weight: 1 is the default.

2. Pretrain

{"text": "Raw corpus document here..."}

Used for continued pretraining rather than supervised fine-tuning. One document per line.

3. Function-calling Instruct

Mistral enforces an exact-9-character ID rule for id and tool_call_id: must match the regex [A-Za-z0-9]{9}. Top-level tools is required. Arguments must be a JSON string, not an object.

{"messages":[
  {"role":"system","content":"Use tools when helpful."},
  {"role":"user","content":"Weather in Paris?"},
  {"role":"assistant","tool_calls":[{
    "id":"aB3kL9mNz",
    "type":"function",
    "function":{"name":"get_weather","arguments":"{\"city\":\"Paris\"}"}
  }]},
  {"role":"tool","tool_call_id":"aB3kL9mNz","name":"get_weather","content":"{\"temp_c\":17}"},
  {"role":"assistant","content":"It's 17 C in Paris."}
],
"tools":[{"type":"function","function":{
  "name":"get_weather",
  "description":"Get current weather",
  "parameters":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}
}}]}

4. Pixtral / vision content

For Pixtral fine-tuning, images live inside a typed content array. Base64 data URIs must be prefixed with data:image/jpeg;base64, (or the corresponding MIME).

{"messages":[
  {"role":"user","content":[
    {"type":"text","text":"What's in this image?"},
    {"type":"image_url","image_url":"data:image/jpeg;base64,/9j/4AAQ..."}
  ]},
  {"role":"assistant","content":"A red barn in a wheat field."}
]}

Mistral fine-tune limits (from validate_data.py)

Note the contradiction with inference: training data must end on assistant, but at inference time the last role must be user or tool. The error Expected last role User or Tool ... but got assistant only fires at serve time, not training.

What this validator checks

• Each line is valid JSON, one record per line.
• Top-level messages is present and is a non-empty array (or text for pretrain mode).
• Every role is one of system, user, assistant, tool. No developer, no human/bot.
• At least one assistant turn; last message must be assistant in training data.
• Non-empty content (or valid tool_calls if content is null/missing).
• tool_calls shape: 9-character id, type: "function", function.name, function.arguments as a JSON string.
• tool role messages have a matching 9-character tool_call_id.
• Assistant turn cannot have both content and tool_calls populated (the InvalidAssistantMessageException rule).
• File size and conversation count flagged against the 1,000-byte / 10-conversation minima.

Common mistakes this validator catches

{"messages":[{"role":"user","content":"Hi"}]}
// Error: missing assistant turn — model has no target to learn.

{"messages":[{"role":"developer","content":"..."}]}
// Error: Mistral does not support 'developer' role (use 'system').

{"messages":[
  {"role":"assistant","tool_calls":[{"function":{"name":"f"}}]}
]}
// Error: tool_calls missing 'id', 'type', and 9-char ID format.

{"messages":[
  {"role":"assistant","content":"hello","tool_calls":[...]}
]}
// Error: InvalidAssistantMessageException — content AND tool_calls together.

{"messages":[
  {"role":"user","content":"Hi"},
  {"role":"assistant","tool_calls":[{"id":"call_abc","type":"function",...}]}
]}
// Error: tool_call id 'call_abc' is 8 chars — must be exactly 9.

Real error strings from La Plateforme and validate_data.py

How to use it

1. Drop a .jsonl file or paste records directly.
2. Click Validate. Each line is parsed and checked against the same rules validate_data.py enforces.
3. Inspect the Error List — every issue maps to either a La Plateforme upload error or a validate_data.py exception name.
4. Download valid examples only rebuilds a clean file with broken lines stripped.
5. Upload to Mistral AI Studio or run mistral-finetune locally.

Tips & common pitfalls

• The 9-character tool-call ID is non-negotiable. OpenAI's call_abc123-style IDs will fail Mistral validation. Use a generator that emits exactly 9 alphanumeric characters.
• Arguments are strings, not objects. "arguments": "{\"city\":\"Paris\"}" ✓ — "arguments": {"city": "Paris"} ✗.
• Content XOR tool_calls on assistant turns. Never both, never neither.
• The deprecated docs banner is real. Mistral's docs.mistral.ai/guides/finetuning/ page has a deprecation banner; fine-tuning now lives on Mistral AI Studio and Azure AI Foundry.
• Pixtral vs v3 tokenizer. Self-hosted mistral-finetune only supports v3-tokenizer weights. Pixtral uses the Tekken tokenizer; train it on Mistral AI Studio, not the self-hosted repo.
• Base64 needs the data URI prefix. Pixtral expects data:image/jpeg;base64,..., not raw base64.

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.
• Llama Fine-Tune ValidatorChatML, ShareGPT, Alpaca.
• Fine-tune dataset editorVisual editor; export to any provider format.
• Token counterPer-line token estimate for cost planning.