Upload or build a dataset

What it does

Datasets are collections of (input, output) pairs used for training. You can upload a JSON or CSV file directly, or build one from feedback collected on past chat completions. The Datasets page lists all datasets, shows row counts, and lets you view, export, and delete them.

Prerequisites

• For upload: a JSON or CSV file with input and output columns.
• For build-from-feedback: at least one chat completion with feedback (guide 03).

Steps

1. Open Datasets page

Click Datasets in the left rail. The page lists existing datasets with row counts and created-at timestamps.

2a. Upload a dataset

Click the Upload button (top-right). A file picker opens.

• Select your JSON or CSV file (one row per line in JSON; standard CSV headers + rows).
• The file is POSTed to POST /admin/datasets.
• A success toast shows the assigned dataset_id.

JSON format example:

{"input": "What is 2 + 2?", "output": "4"}
{"input": "What is the capital of France?", "output": "Paris"}

2b. Build from feedback

Click Build from Feedback (if it appears). A form opens with:

• Base dataset (optional) — start with existing rows.
• Filter by model (optional) — narrow to feedback on a specific model.
• Strategy — DPO (preferred) or SFT.
• Balanced negatives — true/false (default true; balances up/down ratings).

Click Build. The job runs in the background; you'll see a new dataset appear in the list once complete.

3. View dataset rows

Click a dataset name to open the DatasetViewer. The drawer shows:

• Row count — total pairs.
• Columns — input, output (and metadata if present).
• Paginated list — click a row to expand and view full text.

4. Export to argilla format

Click the Export button in the drawer. The dataset is downloaded as a .jsonl file compatible with the Argilla annotation tool.

(Direct API) Upload via curl

curl -s -X POST \
  -H "Authorization: Bearer $BIGWAVE_TOKEN" \
  -F "file=@your-dataset.json" \
  "http://localhost:8080/admin/datasets" | jq '.dataset_id'

Returns the new dataset_id.

Verify

• After upload, your dataset appears in the Datasets list.
• Click it; the DatasetViewer opens and shows the row count.
• Export to argilla; a .jsonl file downloads.
• Verify the file format:

  head -1 downloaded-file.jsonl | jq

Common issues

Symptom	Cause	Fix
"400 Bad Request" on upload	File format is invalid (malformed JSON or CSV without required columns)	Check your file has input and output fields; use jq . < file.json to validate JSON
"No datasets to build from"	No feedback exists yet or all feedback is on deleted completions	Send and rate some chats first (guides 01 & 03)
Export returns empty file	Dataset has 0 rows	Build or upload a non-empty dataset first
Build-from-feedback hangs	Job is still running	Refresh the page after 10s; status updates from the backend

Related

• Capture user feedback (👍/👎) — generates feedback for build-from-feedback.
• Train a LoRA adapter (DPO) — datasets are the input to training jobs.
• Run an eval against a golden set — evals also use datasets as fixtures.