Generate training data for your model.
Turn your tasks into a private dataset for post-training. Use batch inference on your own prompts in OpenAI Batch JSONL with a pinned public model, then inspect and export the verified responses.
- Input
- Your prompts as OpenAI Batch JSONL
- Model
- Pinned public Hugging Face revision
- Sampling
- Set per request
- Output
- Completion and error files
An example configuration, not an active order or a delivered dataset.
Your training data. Ready to inspect and use.
Inspect and export completions matched to the identifiers in your input. A separate error file explains which requests could not be completed.
- Completion file
- Per-request errors
- Batch status
From a task to a dataset.
- 01
Choose the task
Write one chat-completion or completion request per JSONL line. Give each request a unique custom ID.
- 02
Set the scope
Choose a supported public model at an exact revision and set the sampling and output-token limits. Submit the batch with an authorized API key.
- 03
Inspect the delivery
Follow the batch status. Download the completion and error files, then join them to your inputs by custom ID.
promptstringImplement deduplicate(items) preserving order.
responsestringdef deduplicate(items):
seen = dict.fromkeys(items)
return list(seen)View JSONL row
{"prompt":"Implement deduplicate(items) preserving order.","response":"def deduplicate(items):\n seen = dict.fromkeys(items)\n return list(seen)"}Built around your workload.
Eligibility and scope depend on the model, token limits and batch size. The Batch API guide describes how the service estimates charges and reconciles verified output.
Read the Batch API guide Try a sample JSONL batch Compare Batch and evaluation workflows- Model identity
- Choose a qualified public Hugging Face model and pin its commit.
- Input
- Provide your prompts as OpenAI-compatible JSONL.
- Generation controls
- Set output-token limits and sampling. Thinking mode is available where supported.
Start with a representative task.
Batch execution requires approved dataset access and an API key with orders:write. Dataset generation from the catalog or an environment is planned.