Datasets

A dataset is the input to a fine-tuning job. You create one once and reuse it across as many jobs as you want — every job starts from a validated dataset, so you never have to re-upload to retry training with different hyperparameters.

Datasets live under Compute → Datasets in the console. That is the working list: create, validate, and train.

Datasets shared across accounts live somewhere else — Registry → Datasets. See Dataset registry for browsing what others have published, and for publishing your own.

Two upload modes

When you click New Dataset, you pick one of two paths.

Upload (chat format)

Use this if you already have training examples in OpenAI's chat-completion format. Each line of the JSONL file is one training example. The validator runs on upload and your dataset moves directly to ready to train.

Seeds + skill prompt

Use this if you have prompts but no answers — the common case for distilling a skill from an existing teacher model. Each line of the JSONL file is a seed (a user message), and you also paste a skill prompt that defines the persona/tone/format. Your dataset lands at ready to synthesize, and from the dataset detail page you trigger synthesis: pick a teacher model (DeepSeek R1, Qwen 3, etc.), paste your teacher API key (BYOK), and EcoLink fans each seed through the teacher to produce a chat-format child dataset at ready to train.

The teacher-gen path is the bulk of how managed fine-tuning gets used — you typically don't have hand-curated examples for a new skill, but you do have a sense of the prompts users will send. See Synthesis for the full flow including teacher choices, BYOK key handling, and rejection sampling.

Chat-format JSONL

Each line is one JSON object with a messages array. Two equivalent shapes are accepted.

Preferred — assistant turn inside messages[]

{"messages":[{"role":"system","content":"You are a marketing copywriter for a SaaS analytics product."},{"role":"user","content":"Write a one-sentence tagline that emphasizes speed."},{"role":"assistant","content":"Ship insights at the speed of your data — analytics that keeps up."}]}
{"messages":[{"role":"system","content":"You are a customer-support agent. Be concise and reassuring."},{"role":"user","content":"My export download keeps failing at 80%."},{"role":"assistant","content":"That usually means the export hit a 5-minute timeout. Try the 'Email me when ready' option in the export dialog — it works in the background and you'll get a download link when it's done. If it still fails, send me the export ID."}]}

Legacy — top-level completion field

{"messages":[{"role":"system","content":"You translate English to formal Japanese."},{"role":"user","content":"Could you send me the report by tomorrow?"}],"completion":"明日までにレポートをお送りいただけますでしょうか。"}

The validator reshapes this into the first form before training. Both work; pick whichever matches the data you already have.

What the validator requires

• One JSON object per line, no trailing commas.
• messages is a non-empty array; each entry has role (system | user | assistant) and a non-empty content.
• The conversation must end with an assistant turn — either as the last messages[] entry, or via a top-level completion field. This is the training target, so without it there is nothing to learn.
• At least 10 examples.
• Each line is at most 1 MB, and the whole file at most 100 MB.

A system message is not required. Many good datasets have one on every row and some have none — if you want the model to absorb a particular voice, include it consistently rather than on some rows only.

Warnings (the dataset is still usable, but the console says so):

• Fewer than 1 000 examples — a narrow task usually needs more than this to converge to something stable.

Row length is not checked against your training settings

Validation does not compare your rows to the sequence length a job will use. Training truncates anything longer than max_seq_length (4 096 by default), so a very long row is silently cut — and if its assistant answer falls past the cut, that row teaches nothing.

The Structure tab on the dataset page reports your longest row and warns when it exceeds the default. Check it before launching a job on long conversations.

Seeds JSONL

For the teacher-gen path. Each line is one user message that the teacher model will answer.

{"input":"Write a one-sentence tagline for a SaaS analytics tool that emphasizes speed."}
{"input":"Write a 3-bullet announcement post for a new dashboard feature called 'Smart Filters'."}
{"input":"Suggest five subject-line variations for a winback email to lapsed users."}

You also provide a skill prompt in the wizard — the system prompt that gets pasted as messages[0] of every generated row. Three things to nail in a skill prompt: persona (who's answering), voice + tone (how do they write), and format constraints (length, list vs paragraph, forbidden phrases). Example:

You are a marketing copywriter for a B2B SaaS analytics product called Pulse. You write in a confident, direct voice — short sentences, concrete benefits, no hype words like "revolutionary" or "game-changing". Outputs default to one paragraph unless the prompt asks for a list. Always center the customer's outcome (faster decisions, fewer surprises) over feature names.

The wizard ships three preset skill prompts (Marketing copywriter, Customer-support agent, Code reviewer) you can drop in and edit from a working baseline.

Validation lifecycle

1. You click Create & Upload. Your file uploads straight from the browser, so a large dataset does not have to pass through the console.
2. Validation starts automatically, usually within seconds.
3. Every row is checked against the rules above, and the dataset moves to ready to train, ready to synthesize, or invalid.
4. Validation also writes a canonical copy of your data — the exact rows a training job will read — and records what shape it found: example count, tokens per row, conversation turns, and how the roles are distributed. That summary is what the Structure tab shows.

If a dataset is invalid, the detail page shows the first ~50 line errors with line numbers. Fix in your source, re-upload as a new dataset.

The dataset page

Open any dataset to find four tabs alongside its status.

Card — what this dataset is, in prose. A short summary plus a longer write-up covering where the data came from, how it was cleaned, what it is not suitable for, and how it is licensed. If you are the owner you can edit it here; the editor asks one question at a time rather than handing you an empty box.

Viewer — the actual rows, twenty at a time, laid out message by message. It shows the canonical copy written during validation, so what you see is exactly what a training job reads. Paging stops after the first 500 rows; use Download to inspect the whole file.

Structure — the fields each row must have, and what validation measured: how many examples, tokens per row (average and longest), conversation turns, and how many rows carry a system message. This is where the long-row warning appears.

Usage — a link that opens the fine-tuning wizard with this dataset already selected, a ready-to-run API call, and the row format if you want to produce more data in the same shape.

Older datasets show no statistics — they were validated before this existed. Re-uploading computes them.

Reusing a dataset

The same dataset can power any number of fine-tuning jobs. Pick it from the wizard's dataset dropdown each time. Useful for:

• Trying multiple base models against the same data.
• Sweeping hyperparameters (LoRA rank, learning rate, epochs).
• Re-running with a fresh seed.

The dataset row shows used in N jobs so you don't accidentally delete one that's still being trained against. The Delete button is disabled while any active job references it.

Parent / child datasets

When you trigger synthesis on a seeds dataset, EcoLink creates a child dataset at valid_seeds → synthesizing → ready_for_training. The child references its parent for audit. You can train against the child as many times as you want; the parent stays available for re-synthesis if you want different teacher settings.

parent (seeds, skill prompt)
   └─ child (chat format, ready_for_training)   ← train against this

Open the child to see its from #<parent_id> link.

Limits

• File size: 50 MB per upload. Larger files are rejected at the gateway with HTTP 413.
• At least 100 examples to start training (1 000+ recommended for usable quality).
• Cost: $0 for upload, $0 for validation. Synthesis is also free on EcoLink — you pay your teacher provider directly via BYOK.

If you have a dataset larger than 50 MB, split it: train against a 50 MB shard first to validate the recipe, then concatenate before re-running. Or contact support for a raised limit on your account.

Next steps

• Synthesis — generate chat-format data from seeds + a teacher model
• Launching a job — train an adapter against a ready dataset