Datasets

Run BiOS supports multiple dataset formats and provides built-in validation, and preview so your data is ready for training.

Supported Formats

Run BiOS accepts one file type, .jsonl (one JSON object per line), and exactly one shape per training method. This is the industry-standard chat-transcript convention, so datasets prepared for other major fine-tuning services work here unchanged.

{"messages": [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "What is the capital of France?"}, {"role": "assistant", "content": "Paris."}]}

{"text": "One document of plain text per line."}

Files in other layouts (instruction/output, prompt/completion, question/answer, ShareGPT conversations, preference pairs) and other containers (CSV, Parquet, JSON arrays, plain text files) are refused at upload with a message naming what was detected and the exact shape to convert to.

Uploading Datasets

Uploading a dataset is the first step in any training workflow. The platform handles validation, format detection, and preview generation automatically.

1. Navigate to Datasets in the sidebar. You will see your existing datasets listed with their name, format, row count, size, source, and upload date.
2. Click Upload Dataset in the top right corner. A drop zone appears where you can drag and drop your file or click to browse your filesystem.
3. Select your file. The platform accepts files up to 200MB. A progress bar shows the upload status in real time.
4. Once uploaded, the platform automatically:
   • Detects the file format (JSONL, Parquet, or CSV)
   • Counts the total number of rows
   • Identifies all columns and their data types
   • Validates the structure for common issues (missing fields, malformed JSON, encoding errors)
5. A green Dataset is valid and ready for training banner confirms the file passed all validation checks.
6. If the file is refused, the message names the detected layout and the exact shape to convert to (see below).

You can also import datasets from HuggingFace Hub using the Import from HF button. This opens a search interface where you can browse public datasets or access your private repos (after connecting your HuggingFace account under Integrations).

Format Validation

Before upload completes, the platform validates your dataset for common issues: missing required fields, inconsistent schemas across rows, encoding errors, and empty records. You can also use the standalone validation endpoint to check a file before uploading.

Converting Other Layouts

The platform does not remap columns. If your data is in another layout, convert it to the canonical shape before uploading. For example, an instruction/output row becomes a messages array with one user turn (the instruction, plus any input) and one assistant turn (the output). ShareGPT from/value turns map to role/content with human → user and gpt → assistant.

HuggingFace Import

You can import datasets directly from the HuggingFace Hub:

1. Public datasets: use the Hub Search to find and preview any public dataset, then import it with one click.
2. Private datasets: first connect your HuggingFace account under Integrations by adding your API token. Then browse your private repos and import from them.

Preview and Display

Each dataset has two preview modes: a paginated raw preview showing all columns and rows (supports pagination), and a display previewthat renders formatted instruction/response pairs for quick inspection.