Skip to main content
--rollout-data-yaml-config replaces --prompt-data when a run trains on several datasets at once, for example math, science QA and agentic tasks. Miles draws each rollout step from every dataset in a fixed ratio, so there is no need to merge and shuffle the files by hand.
The flag also accepts an inline base64:<payload>, like --eval-config.

​
Fields

Unknown fields are rejected, so a typo fails at startup instead of being ignored. A custom_generate_function_path or custom_agent_function_path that declares its own options (config_class or add_arguments) gets them registered as flags, just like one passed through --custom-generate-function-path, so two agentic datasets can each run their own generate function and agent. These flags live in one run-wide namespace: two hooks that declare the same flag identically share it, so every dataset using those hooks sees the same value. A hook that re-declares only some of another hook’s flags, or declares one differently, is rejected at startup. To give two datasets different values for an option, give their hooks distinct flags (or put the value in the dataset’s agent block). custom_agent_function_path and max_seq_len are read only by miles.rollout.generate_hub.agentic_tool_call.generate, or by a wrapper around it that sets reads_sample_agent_fields = True on itself. A dataset that sets them under any other generate function, or an agentic dataset with neither custom_agent_function_path nor --custom-agent-function-path, is rejected at startup. Eval samples do not come from the mixer, so an agentic eval still needs --custom-agent-function-path. custom_rm_path and rm_type become the samples’ reward_spec, so they follow the per-sample reward rules in Customization, checked at startup: a function with its own config must be the --custom-rm-path one, and under --group-rm a batched --custom-rm-path scores every group, so a dataset that sets rm_type or names a different function is rejected. A custom reward function also outranks any rm_type, so with --custom-rm-path set, a dataset’s rm_type needs its own custom_rm_path too.

​
How prompts are drawn

  • Weights, not counts. samples_per_step values need not add up to --rollout-batch-size. Every request is split by weight, and the remainder carries over to the next request, so the long-run ratio is exact even when fully async rollout asks for one prompt at a time.
  • Mixed by prompt. Datasets are mixed before each prompt is copied --n-samples-per-prompt times, so every group still holds the responses to a single prompt.
  • Separate epochs. Each dataset has its own position, epoch and shuffle, and wraps around on its own. The checkpoint saves all of them, and resuming also re-checks every file’s SHA-256.

​
Resuming with different datasets

A resumed run continues every dataset from its saved position, so --load refuses a YAML that lists other datasets, a changed file, a dataset filtered shorter than its saved position, or any other changed dataset field (groups drawn before the save still carry the old generate function, agent or reward). Changing only the samples_per_step weights is fine. To switch datasets mid-training, add --rollout-data-reset-on-load: the model, optimizer, rollout step and sample ids still resume from --load, while every dataset starts from its beginning and the prompt groups and untrained batches drawn from the old datasets are dropped. The reset happens only when the checkpoint does not match the current YAML, so the flag can stay in the launch command: later resumes from the run’s own checkpoints continue in place. The same flag moves a run started with --prompt-data onto a mix. The other way needs no flag: a mix resumed under --prompt-data starts that dataset from its first prompt, with the same drops and the same sample ids. Every sample carries metadata["dataset_name"], and the rollout metrics add rollout/dataset/<name>/{episode_raw_reward,response_len/mean,truncated_ratio,num_training_samples}.

​
Limits

  • Set --num-rollout: an epoch has no single length across datasets that wrap independently, so --num-epoch is not supported.
  • --partial-rollout and a custom --data-source-path are not supported.
  • The ratio holds for the prompts drawn. When dynamic sampling filters out groups, the refill is drawn by the same weights, not from the dataset whose groups were dropped.