Troubleshooting¶
Locate problems across data, training, deployment, and inference to restore your workflow.
Identify the stage first, then inspect the resource's status and error details. Keep the job, dataset version, or deployment ID so a request can be connected to its execution.
Data cannot be used for training¶
Check validation status and valid sample count in My data. SFT input uses messages JSONL with one independent JSON object per line. Fix structural errors before uploading again. upload_expired means the upload did not complete.
A dataset row alone does not establish a successful upload. See Data for the format.
Training cannot be submitted or is queued¶
The training summary identifies missing models, datasets, or invalid configuration. Resolve local checks and the training precheck before submitting.
Serverless jobs can wait for available resources. Queued and failed are distinct states; inspect the existing job in My training rather than submitting duplicates. For a failed job, inspect its detailed errors and logs.
Training succeeded but the model is unavailable¶
Confirm the version and required artifacts in My models, then create a deployment. Verify loading, health, and routing before selecting the service on the API access page.
API requests fail¶
| Symptom | Check first |
|---|---|
| 401 / 403 | Key validity and permission to access the target service |
| Model missing or unavailable | API access Model ID and deployment readiness |
| Path not found | Matching environment and Base URL; duplicated path segments |
| Timeout or unavailable service | Deployment health, routing, and client timeout settings |
Reproduce with the minimal API access example. Confirm one request before investigating application concurrency, context length, or extra parameters.
Evaluation exits abnormally or has no score¶
Check status and trial progress in My evaluations. Confirm model access and a benchmark compatible with the selected evaluation type. Execution completion, task success, and score availability are different outcomes. Missing scores should not be treated as zero scores or successful evaluations.
Request help¶
Include the time, reproduction steps, job or deployment ID, status, and redacted errors. For data issues, provide a minimal example. Do not send API keys, passwords, or sensitive datasets.
Contact sales@mindlab.ltd to reach the team for routing your request.