Usage
Arguments
string
required
New Training name.
Options
string
default:"\"sft\""
Training method: sft or gisting.
string
Catalog Base Model source; pass exactly one of
--base-model or --model-version.string
Exact Model Version source; pass exactly one of
--base-model or --model-version.string
Dataset to train on: fills the empty cells of its split column (a new split column, 80/10/10, when it has none) and freezes a slice named after the Training; pass
--dataset or --slice.string
Ready Dataset Slice that freezes a split column; pass
--dataset or --slice.string
Split column to train with, when the dataset has more than one.
string
Train parts to learn on, comma-separated (default: every part with the train role).
integer
SFT LoRA adapter rank: 8, 16, or 32 (default 16).
string
Gisting soft tokens: 8, 32, or 64, or a fraction of the prompt “1/4”, “1/8”, “1/16”, “1/32” (default “1/8”).
integer
Training epochs, 1 through 5 (default 3).
number
Optimizer learning rate (default 0.00015 for SFT, 0.03 for gisting).
integer
Sequences per optimizer step: 8, 16, 32, or 64 (default 8).
integer
Token budget per example, 64 through 16384 (default 12288 for SFT, 3072 for gisting).
string
Required for gisting: non-empty UTF-8 prompt text file, at most 256 KiB.
boolean
default:"false"
Print JSON.
Inherited options
boolean
Output the current version.
string
Project root; otherwise search the working directory and parents; init and report use only the working directory.
string
API key override; login stores the key it verifies.Environment fallback:
TOLOKA_SCOPED_TOKEN, TOLOKA_API_KEY.string
Platform base URL override; login saves the selected URL.Environment fallback:
TOLOKA_BASE_URL.string
Target project; overrides toloka.json.Environment fallback:
TOLOKA_PROJECT_ID.