akka optimize training distillations create

Start a distillation

Synopsis

Start a distillation. The teacher answers every decision in the source dataset. Optimize then registers the accepted answers as a version of the target dataset. The distillation completes.

The teacher first answers the smoke decisions on one GPU. The smoke takes the first 16 decisions of the source, and all but one must be accepted. Use --smoke to specify the decisions instead. Every specified decision must be accepted. Once the smoke passes, the teacher answers decisions across the whole source on --replicas replicas, with one GPU each, until the distillation ends.

References in --teacher, --from and --to without a namespace use the training project’s namespace.

Omitted settings use these defaults: --max-tokens 4096, --max-model-length 65536, --temperature 1.0, --top-p 1.0, --seed 0, --thinking off and --replicas 1.

akka optimize training distillations create --project PROJECT --teacher MODEL --from DATASET --to NAME [flags]

Examples

  # Start a distillation whose smoke takes the first decisions of the source
  akka optimize training distillations create --project support-triage \
    --teacher google/gemma-4-31b-it:v1 --from triage/conversations:v2 --to triage-distilled

  # Name the smoke's decisions
  akka optimize training distillations create --project support-triage \
    --teacher google/gemma-4-31b-it:v1 --from triage/conversations:v2 --to triage-distilled \
    --smoke 0:1,1:1,2:3

  # Answer on four replicas, with reasoning, and wait for it
  akka optimize training distillations create --project support-triage \
    --teacher google/gemma-4-31b-it:v1 --from triage/conversations:v2 --to triage-distilled \
    --smoke 0:1,1:1 --replicas 4 --thinking on --wait

Options

      --description string     Description for the distillation
      --exit-status            Exit with a non-zero status if the distillation fails or is cancelled. Requires --wait
      --from string            Source chat dataset whose assistant messages are the decisions, as NS/NAME[:VERSION]
  -h, --help                   help for create
      --instructions string    Text appended to each prompt's system message
      --max-model-length int   Maximum combined token count for a prompt and its answer (default 65536)
      --max-tokens int         Maximum number of tokens the teacher generates for one decision (default 4096)
      --project string         Training project name or ID whose GPU quota the distillation counts against
      --replicas int           Teacher replicas after the smoke, one GPU each (default 1)
      --seed int               Sampling seed
      --smoke string           Decisions the smoke answers on one GPU, as ROW:MESSAGE pairs separated by commas, such as 0:2,5:1. If omitted, the smoke uses the first 16 decisions of the source
      --teacher string         Model that answers, as a registry reference NS/NAME[:VERSION]: a base model or an adapter over one
      --temperature float      Sampling temperature (default 1)
      --thinking string        Whether the prompt asks the teacher to reason first: on or off (default "off")
      --to string              Named dataset for the accepted rows, as NS/NAME
      --top-p float            Nucleus sampling threshold (default 1)
      --wait                   Wait until the distillation ends

Options inherited from parent commands

      --cache-file string    location of cache file (default "~/.akka/cache.yaml")
      --config string        location of config file (default "~/.akka/config.yaml")
      --context string       configuration context to use
      --disable-prompt       Disable all interactive prompts when running akka commands. If input is required, defaults will be used, or an error will be raised.
                             This is equivalent to setting the environment variable AKKA_DISABLE_PROMPTS to true.
      --force                Skip the confirmation prompt
      --jq string            Filter the JSON output with a jq expression
      --json string          Output JSON with only these fields; --json= lists the fields a command offers
      --no-color             Print without color or a redrawn progress line, as NO_COLOR does
  -o, --output string        set output format to one of [text,json,json-compact,go-template=] (default "text")
      --page-mode string     the mode for paging, either paged, buffered or auto. (default "auto")
  -q, --quiet                set quiet output (helpful when used as part of a script)
      --region string        Region to use if the project has more than one region
      --timeout duration     client command timeout (default 10s)
      --trainer-url string   Trainer service URL (default TRAINER_URL, then AKKA_OPTIMIZE_URL, then a local dev-mode service, then http://localhost:9001)
      --use-grpc-web         use grpc-web when talking to Akka APIs. This is useful when behind corporate firewalls that decrypt traffic but don't support HTTP/2.
      --verbose              set verbose output

SEE ALSO