> ## Documentation Index
> Fetch the complete documentation index at: https://fireworks.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# firectl rlor-trainer-job create

> Reference for the firectl rlor-trainer-job create command: launch a Fireworks RL-on-rollouts trainer job from the CLI, with recipe, dataset, and model flags.

```
firectl rlor-trainer-job create [flags]
```

### Examples

```
firectl rlor-trainer-job create \
	--base-model llama-v3-8b-instruct \
	--dataset sample-dataset \
	--output-model name-of-the-trained-model
	--rewards key1=0.5,key2=0.5

```

### Flags

```
      --base-model string                         The base model for the rlor trainer job. Only one of base-model or warm-start-from should be specified.
      --dataset string                            The dataset for the rlor trainer job. (Required)
      --output-model string                       The output model for the rlor trainer job.
      --job-id string                             The ID of the rlor trainer job. If not set, it will be autogenerated.
      --warm-start-from string                    The model to warm start from. If set, base-model must not be set.
      --evaluation-dataset string                 The evaluation dataset for the fine-tuning job.
      --epochs int32                              The number of epochs for the fine-tuning job.
      --learning-rate float32                     The learning rate for the fine-tuning job.
      --max-context-length int32                  Maximum token length per sequence. When unset, defaults to the validated training shape's max supported context length. Capped (not pinned) by the shape: smaller user values are honored.
      --batch-size int32                          Legacy V1 token budget. On V1 RLOR paths, limits packed tokens per microbatch and remains active when --batch-size-samples is set. Training V2 rejects nonzero values.
      --batch-size-samples int32                  Sample-count batching control. On V1 RLOR paths it limits samples per gradient batch and works alongside --batch-size; on Training V2 SFT and DPO paths it maps to examples or preference pairs per optimizer step.
      --gradient-accumulation-steps int32         The number of batches to accumulate gradients before updating the model parameters. The effective batch size will be batch-size multiplied by this value. (default 1)
      --learning-rate-warmup-steps int32          The number of learning rate warmup steps for the rlor trainer job.
      --lora-rank int32                           The rank of the LoRA layers for the fine-tuning job.
                                                   (default 8)
      --wandb-api-key string                      [WANDB_API_KEY] WandB API Key. (Required if any WandB flag is set)
      --wandb-project string                      [WANDB_PROJECT] WandB Project. (Required if any WandB flag is set)
      --wandb-entity string                       [WANDB_ENTITY] WandB Entity. (Required if any WandB flag is set)
      --wandb                                     Enable WandB
                                                  
                                                  
      --aws-credentials-secret string             [AWS_CREDENTIALS_SECRET] AWS credentials secret (mutually exclusive with --aws-iam-role)
      --aws-iam-role string                       [AWS_IAM_ROLE_ARN] AWS IAM role ARN (mutually exclusive with --aws-credentials-secret)
                                                  
                                                  
      --azure-credentials-secret string           [AZURE_CREDENTIALS_SECRET] Azure credentials secret
      --azure-managed-identity-client-id string   [AZURE_MANAGED_IDENTITY_CLIENT_ID] Azure managed identity client ID for Workload Identity Federation
      --azure-tenant-id string                    [AZURE_TENANT_ID] Azure tenant ID (required with --azure-managed-identity-client-id)
                                                  
      --rewards strings                           A list of reward weights as key=value pairs (e.g., --rewards key1=1.0,key2=0.5).
      --display-name string                       The display name of the supervised fine-tuning job.
      --quiet                                     If set, only errors will be printed.
      --eval-auto-carveout                        If set, the evaluation dataset will be auto-carved.
      --training-shape string                     Training shape selector for service-mode launches. Accepts training shape names or training shape version names; a shape name uses its latest validated version.
      --trainer-replicas int32                    Data-parallel replica count for service-mode trainers. If not specified, defaults to 1.
      --inactivity-timeout duration               Trainer inactivity timeout (for example, 30m or 2h, maximum 3h). The trainer reports tracked activity, including trainer API operations and active-session heartbeats. If no tracked activity is observed for this duration, the trainer is automatically stopped. If omitted or 0, defaults to 60 minutes. GPU usage continues to accrue while the trainer is running.
      --disable-inactivity-cleanup                Disable trainer inactivity cleanup. Only supported when creating trainers in the fireworks account; GPU usage continues to accrue while the trainer is running.
      --dry-run                                   Print the request proto without running it.
  -o, --output Output                             Set the output format to "text", "json", or "flag". (default text)
  -h, --help                                      help for create
```

### Global flags

```
  -a, --account-id string   The Fireworks account ID. If not specified, reads account_id from ~/.fireworks/auth.ini.
      --api-key string      An API key used to authenticate with Fireworks.
  -p, --profile string      fireworks auth and settings profile to use.
```


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.