Troubleshooting model customization issues

This section summarizes errors that you might encounter and what to check if you do.

Permissions issues

If you encounter an issue with permissions to access an Amazon S3 bucket, check that the following are true:

1. If the Amazon S3 bucket uses a CM-KMS key for Server Side encryption, ensure that the IAM role passed to Amazon Bedrock has kms:Decrypt permissions for the AWS KMS key. For example, see Allow a user to enccrypt and decrypt with any AWS KMS key in a specific AWS account.

2. The Amazon S3 bucket is in the same region as the Amazon Bedrock model customization job.

3. The IAM role trust policy includes the service SP (bedrock.amazonaws.com).

The following messages indicate issues with permissions to access training or validation data in an Amazon S3 bucket:

Could not validate GetObject permissions to access Amazon S3 bucket: training-data-bucket at key train.jsonl
Could not validate GetObject permissions to access Amazon S3 bucket: validation-data-bucket at key validation.jsonl

If you encounter one of the above errors, check that the IAM role passed to the service has s3:GetObject and s3:ListBucket permissions for the training and validation dataset Amazon S3 URIs.

The following message indicates issues with permissions to write the output data in an Amazon S3 bucket:

Amazon S3 perms missing (PutObject): Could not validate PutObject permissions to access S3 bucket: bedrock-output-bucket at key output/.write_access_check_file.tmp

If you encounter the above error, check that the IAM role passed to the service has s3:PutObject permissions for the output data Amazon S3 URI.

Data issues

The following errors are related to issues with the training, validation, or output data files:

Invalid file format

Unable to parse Amazon S3 file: fileName.jsonl. Data files must conform to JSONL format.

If you encounter the above error, check that the following are true:

1. Each line is in JSON.

2. Each JSON has two keys, an input and an output, and each key is a string. For example:

   {
       "input": "this is my input",
       "output": "this is my output"
   }

3. There are no additional new lines or empty lines.

Character quota exceeded

Input size exceeded in file fileName.jsonl for record starting with...

If you encounter an error beginning with the text above, ensure that the number of characters conforms to the character quota in Prepare the datasets.

Token count exceeded

Maximum input token count 4097 exceeds limit of 4096
Maximum output token count 4097 exceeds limit of 4096
Max sum of input and output token length 4097 exceeds total limit of 4096

If you encounter an error similar to the preceeding example, make sure that the number of tokens conforms to the token quota in Prepare the datasets.

Third party license terms and policy issues

The following errors are related to third-party's license terms and their policies:

Fine-tuning materials inconsistent with third-party's license terms

Automated tests flagged this fine-tuning job as including materials that are potentially inconsistent with Anthropic's third-party license terms. Please contact support.

If you encounter the above error, ensure your training dataset does not contain content that is inconsistent with Anthropic's usage policies. If the issue persists, contact AWS Support.

Internal error

Encountered an unexpected error when processing the request, please try again

If you encounter the above error, there might be an issue with the service. Try the job again. If the issue persists, contact AWS Support.