Docs

Dataset format

Yachay accepts JSONL (OpenAI chat / Alpaca / ShareGPT) or CSV / TSV. We auto-detect and normalize to OpenAI chat-style before fine-tuning — drop whatever you have. The submit form lints every row before you pay; if anything’s off, you’ll see exactly which line.

Accepted formats

Yachay auto-detects the dataset format at upload and normalizes everything to OpenAI chat-style JSONL before fine-tuning. You don’t need to convert anything yourself.

Format	Shape	Where it comes from
OpenAI chat JSONL	keys: messages[].role, messages[].content	Default for new fine-tuning
Alpaca JSONL	keys: instruction, input (optional), output	Academic LoRA tutorials, HF datasets
ShareGPT JSONL	keys: conversations[].from, conversations[].value	Older HF datasets, Vicuna-era
CSV / TSV	header row + ≥ 2 columns; first → user, last → assistant	Spreadsheets, exports from databases

Good example

{"messages": [{"role": "system", "content": "You are a polite customer support agent for Acme."}, {"role": "user", "content": "My order #1234 hasn't shipped yet."}, {"role": "assistant", "content": "I'm sorry to hear that. Let me look into order #1234 for you."}]}
{"messages": [{"role": "user", "content": "What's your return policy?"}, {"role": "assistant", "content": "We accept returns within 30 days of purchase, no questions asked. Email returns@acme.com to start the process."}]}
{"messages": [{"role": "user", "content": "Where are you based?"}, {"role": "assistant", "content": "Acme is headquartered in Portland, Oregon, with fulfilment from Indianapolis."}]}

Three examples, one per line. The first has a system prompt; the other two start straight with the user.

Common mistakes

{"messages": [{"role": "user", "content": ""}]}
{"input": "What's the weather?", "output": "Sunny."}
{"messages": [{"role": "user", "content": "Hi"}, {"role": "user", "content": "Hi again"}]}
{"instruction": "Translate to French", "input": "Hello", "output": "Bonjour"}

Line 1: empty content. Rejected.
Line 2: looks like Alpaca but is missing the required instruction field — bare input/outputisn’t a recognized shape and gets dropped at normalization.
Line 3: two consecutive user turns. Rejected.
Line 4: a valid Alpaca row mixed into an OpenAI-chat file. Either format works on its own; mixing in one file makes auto-detection pick one and silently drop the others. Stick to a single format per file.

CSV / TSV — how columns map

First column becomes the user turn.
Last column becomes the assistant turn.
Any columns in between are ignored. A 4-column file like id,prompt,metadata,response uses only id → user and response → assistant — middle columns are silently dropped. If idisn’t what you want as the prompt, reorder so prompt is in the first column.
A header row is required (≥ 2 columns) and is skipped during training. Column namesdon’t matter — only position.
Embedded commas / tabs / newlines inside a quoted field are handled per RFC 4180. A literal " inside a quoted cell is written as "".

Rules

One JSON object per line — no commas, no array wrapper.
The file is JSONL (newline-delimited JSON), not JSON. Most tools that emit "NDJSON" or "line-delimited JSON" produce the right shape.
Each line has a top-level messages array.
Inside, each entry is an object with role and content fields. role is one of system, user, or assistant. content is plain text — no markdown, no images, no nested JSON.
Turns alternate user → assistant.
An optional system message can lead. After that, user and assistant turns must alternate. Two consecutive user turns get rejected.
Examples should be representative.
Yachay trains on what you give it. 50 great examples beat 5000 mediocre ones. Aim for diversity of phrasing and difficulty inside a tight domain.

Limits

Min examples: 20
Recommended: 200–5,000
Max file size: 1 GB
Max content length: Model context window minus 256 tokens

Got something stranger than these five? HuggingFace Parquet, custom JSON schemas, multi-modal — email hello@condorbox.ai with a sample and we’ll send back a converter script.

← All docs · Browse base models