Fine-tuning

Create

To start a new fine-tuning job:

tg fine-tuning create --training-file [FILE_ID | PATH] --model [MODEL]

# Shorthand
tg ft -c --training-file [FILE_ID | PATH] --model [MODEL]

You must provide either --model (to start from a base model) or --from-checkpoint (to resume from a previous job). Before the job is submitted, the CLI prints an estimated price and asks for confirmation. Pass --confirm (or -y) to skip the prompt in scripts and CI.

If --training-file (or --validation-file) is a local path, the CLI uploads the file to the Files API automatically before kicking off the job.

Parameters

Flag	Description
`--training-file/-t [string \| Path]`	required Training file ID from the Files API or a local path to upload. The maximum allowed file size is 25 GB.
`--model [string]`	Base model to fine-tune. See supported models. Required unless `--from-checkpoint` is set.
`--from-checkpoint [string]`	Continue training from a previous fine-tuning job. Format: `JOB_ID/OUTPUT_MODEL_NAME:STEP`. The step is optional; the final checkpoint is used when omitted. Mutually exclusive with `--model`.
`--validation-file/-v [string]`	Validation file ID from the Files API or a local path to upload. Required when `--n-evals > 0`. The maximum allowed file size is 25 GB.
`--suffix [string]`	Up to 40 characters appended to the fine-tuned model name. Recommended to differentiate fine-tuned models.
`--packing/--no-packing`	Whether to use sequence packing for training. Default: enabled.
`--max-seq-length [integer]`	Maximum sequence length to use for training. Required when `--no-packing` is set. Defaults to the maximum allowed for the model and training type.
`--n-epochs/-ne [integer]`	Number of epochs to fine-tune on the dataset. Default: 1. Min: 1. Max: 20.
`--n-evals [integer]`	Number of evaluation loops to run on the validation set. Default: 0. Min: 0. Max: 100.
`--n-checkpoints/-c [integer]`	The number of checkpoints to save during training. Default: 1. One checkpoint is always saved on the last epoch. Must be 1 ≤ n-checkpoints ≤ n-epochs.
`--batch-size/-b [integer \| max]`	Batch size for each training iteration. See supported models for min and max batch sizes per model. Default: `max`.
`--learning-rate/-lr [float]`	Learning rate multiplier. Default: 0.00001. Min: 0.00000001. Max: 0.01.
`--lr-scheduler-type [linear \| cosine]`	Learning rate scheduler type. Default: `cosine`.
`--min-lr-ratio [float]`	Ratio of the final learning rate to the peak learning rate. Default: 0.0. Min: 0.0. Max: 1.0.
`--scheduler-num-cycles [float]`	Number or fraction of cycles for the cosine learning rate scheduler. Must be non-negative. Default: 0.5.
`--warmup-ratio [float]`	Fraction of steps at the start of training to linearly warm up the learning rate. Default: 0.0. Min: 0.0. Max: 1.0.
`--max-grad-norm [float]`	Max gradient norm for gradient clipping. Set to 0 to disable. Default: 1.0. Min: 0.0.
`--weight-decay [float]`	Weight decay for the optimizer. Default: 0.0. Min: 0.0.
`--random-seed [integer]`	Random seed for reproducible training. Uses the server default if unset.
`--confirm/-y`	Skip the price-confirmation prompt. Useful in scripts and CI.
`--train-on-inputs [true \| false \| auto]`	Whether to mask user messages in conversational data or prompts in instruction data. `auto` infers from the data format: Datasets with the `"text"` field (general format): inputs are not masked. Datasets with the `"messages"` field (conversational format) or `"prompt"` and `"completion"` fields (instruction format): inputs are masked. Default: `auto`.
`--train-vision/--no-train-vision`	Update the vision encoder parameters. Default: `false`. Only available for vision-language models.
`--from-hf-model [string]`	Hugging Face Hub repository to start training from. Should match the base model’s architecture and size. When `--lora` is set with `--lora-trainable-modules all-linear`, the modules `k_proj, o_proj, q_proj, v_proj` are targeted for adapter training.
`--hf-model-revision [string]`	Revision (branch name or commit hash) of the Hugging Face Hub model.
`--hf-api-token [string]`	Hugging Face API token for downloading from a private repo or uploading the output model.
`--hf-output-repo-name [string]`	Hugging Face repo to upload the fine-tuned model to.

Weights & Biases

Flag	Description
`--wandb-api-key [string]`	Your Weights & Biases API key. Falls back to the `WANDB_API_KEY` environment variable.
`--wandb-base-url [string]`	Base URL of a dedicated Weights & Biases instance. Leave empty if you are not using a self-hosted instance.
`--wandb-project-name [string]`	Weights & Biases project for your run. Defaults to `together`.
`--wandb-name [string]`	Weights & Biases run name.
`--wandb-entity [string]`	Weights & Biases entity (team or user).

LoRA

Flag	Description
`--lora/--no-lora`	Force LoRA fine-tuning (`--lora`) or full fine-tuning (`--no-lora`). When omitted, the API auto-detects: it defaults to LoRA on most base models, and inherits the parent job’s training type when `--from-checkpoint` is set.
`--lora-r [integer]`	Rank for LoRA adapter weights. Default: 8. Min: 1. Max: 64.
`--lora-alpha [integer]`	Alpha for LoRA adapter training. Default: 8. Min: 1.
`--lora-dropout [float]`	Dropout probability for LoRA layers. Default: 0.0. Min: 0.0. Max: 1.0.
`--lora-trainable-modules [string]`	Comma-separated list of LoRA trainable modules. Default: `all-linear`. See supported modules for LoRA training.

Preference fine-tuning (DPO, RPO, SimPO)

Flag	Description
`--training-method [sft \| dpo]`	Training method. `sft` is supervised fine-tuning; `dpo` is Direct Preference Optimization. Default: `sft`. The DPO method also accepts the RPO and SimPO loss modifiers below.
`--dpo-beta [float]`	Beta parameter for DPO training. Only used when `--training-method dpo`.
`--dpo-normalize-logratios-by-length`	Normalize logratios by sample length. Only used when `--training-method dpo`. Default: `false`.
`--rpo-alpha [float]`	RPO alpha parameter (adds NLL term to the DPO loss). Only used when `--training-method dpo`.
`--simpo-gamma [float]`	SimPO gamma parameter. Only used when `--training-method dpo`.

The id field in the JSON response contains the fine-tune job ID (ft-…) that you use to retrieve status, list events, cancel the job, and download weights.

List

To list past and running fine-tune jobs:

tg fine-tuning list

# Shorthand
tg ft ls

Jobs are listed newest first.

Retrieve

To retrieve metadata for a job, including its current status:

tg fine-tuning retrieve [FT_ID]

List events

To list events of a past or running job:

tg fine-tuning list-events [FT_ID]

Cancel

To cancel a running job:

tg fine-tuning cancel [FT_ID]

List checkpoints

To list saved checkpoints of a job:

tg fine-tuning list-checkpoints [FT_ID]

Each checkpoint includes step, path, created_at, checkpoint_type, and checkpoint (the download selector: model or adapter). When the job uploaded the artifact to the Together model registry, the entry also includes object_id and object_revision_id (for example, ml_… and rv_…). See Model registry object IDs for how these relate to the job-level model_object_id / adapter_object_id fields.

Download model weights

To download the weights of a fine-tuned model, run:

# Download the model to the current working directory.
tg fine-tuning download [FT_ID]

tg fine-tuning download [FT_ID] \
  --output-dir ./models

# Use `tg fine-tuning list-checkpoints` to find the checkpoint index.
tg fine-tuning download [FT_ID] \
  --checkpoint-step 0

tg fine-tuning download [FT_ID] \
  --checkpoint-type adapter

The command downloads Zstandard-compressed (.zst) weights. To extract them, run tar -xf filename.

Parameters

Flag	Description
`--output-dir/-o [Path]`	Output directory.
`--checkpoint-step/-s [integer]`	Download a specific checkpoint’s weights. Defaults to the latest checkpoint.
`--checkpoint-type/-c [merged \| adapter \| default]`	Checkpoint type. `merged` and `adapter` apply to LoRA jobs only; `default` resolves to `merged` for LoRA jobs and to the full model for non-LoRA jobs. Default: `merged`.

Delete

To delete a fine-tuning job:

tg fine-tuning delete [FT_ID]

# Shorthand
tg ft -d [FT_ID]

Parameters

Flag	Description
`--force`	Bypass confirmation prompt.

TOGETHER CLI

COMMANDS

Create

Parameters

Weights & Biases

LoRA

Preference fine-tuning (DPO, RPO, SimPO)

List

Retrieve

List events

Cancel

List checkpoints

Download model weights

Parameters

Delete

Parameters

​Create

​Parameters

​Weights & Biases

​LoRA

​Preference fine-tuning (DPO, RPO, SimPO)

​List

​Retrieve

​List events

​Cancel

​List checkpoints

​Download model weights

​Parameters

​Delete

​Parameters

Create

Parameters

Weights & Biases

LoRA

Preference fine-tuning (DPO, RPO, SimPO)

List

Retrieve

List events

Cancel

List checkpoints

Download model weights

Parameters

Delete

Parameters