veri run — Hero Verb & Project Layout

veri run <config.toml> is a thin shortcut over the explicit veri jobs create / veri deployments create / veri evals create. It reads the required kind field from the config and dispatches. The contract is plural-noun-verb (veri jobs create configs/train.toml); veri run exists to save typing on the common path.

Scaffold a project: `veri init`

veri init my-project
# or, in the current directory:
veri init

veri init NAME writes 9 files into ./NAME. With no NAME, scaffolds into the current directory.

my-project/
├── veri.toml                  # project defaults: gpu_type, gpu_count, base_model
├── reward.py                  # TRL-style reward function stub
├── configs/
│   ├── train.toml             # kind = "train"
│   ├── deploy.toml            # kind = "deploy"
│   └── eval.toml              # kind = "eval"
├── datasets/
│   └── README.md
├── AGENTS.md                  # directory contract for coding agents
├── CLAUDE.md                  # one-liner: "See AGENTS.md."
└── .gitignore

By default init refuses to write into a non-empty directory. Use --force to scaffold alongside existing files (existing files are preserved):

veri init --force

Dispatch rules

Every config starts with a required kind line:

kind = "train"    # or "deploy" or "eval"

veri run <path> reads kind and routes:

`kind =`	Dispatches to
`"train"`	`veri jobs create`
`"deploy"`	`veri deployments create`
`"eval"`	`veri evals create` (coming soon)

Missing kind or an unknown value errors with a “did you mean” hint:

$ veri run nokind.toml
Error: config 'nokind.toml' is missing required `kind` field.
Add one of: kind = "train" | "deploy" | "eval"

Override TOML fields from the CLI

Use --set <dotted.key>=<value> to override any nested value. Convenience long-flags expand to --set internally.

# Submit a training run with two overrides:
veri run configs/train.toml \
  --set model.base=Qwen/Qwen3-8B \
  --set method.learning_rate=2e-6

# Convenience flag (jobs only) — same as --set model.base=...:
veri jobs create configs/train.toml --base-model Qwen/Qwen3-8B

Type coercion

Values parse as TOML scalars: true / false, integers, floats, ISO timestamps, quoted strings. To force a string on a numeric-looking value, use --set-string:

veri jobs create configs/train.toml --set-string job.name=2026-05-12

Repeated flags build a list

veri jobs create configs/train.toml --set tags=experiment --set tags=v2
# Results in tags = ["experiment", "v2"]

Conflict detection

The merger rejects setting a scalar and a sub-key at the same path:

$ veri jobs create train.toml --set model=foo --set model.base=bar
Error: --set conflict: 'model' is set as a scalar; can't also set sub-key 'model.base'

Convenience long-flags

These flags are sugar over --set on veri jobs create:

Flag	Expands to
`--base-model X`	`--set model.base=X`
`--gpu-type X`	`--set resources.gpu_type=X`
`--gpu-count N`	`--set resources.gpu_count=N`
`--dataset X`	`--set dataset.id=X`
`--reward X`	`--set reward.id=X`

Config schemas

`veri.toml` (project root)

[project]
name = "my-project"

[defaults]
gpu_type = "A100-80GB"
gpu_count = 1
base_model = "Qwen/Qwen3-4B"

`configs/train.toml`

kind = "train"

[job]
name = "first-run"
output_model = "tuned-model"

[model]
base = "Qwen/Qwen3-4B"

[dataset]
id = "ds_abc123"

[reward]
id = "rf_abc"
file = "./reward.py"
format = "trl"

[method]
type = "grpo"           # grpo | sft_text | dpo | sft_video_gen (grpo_agentic is being rebuilt; not currently available)
learning_rate = 1e-6
max_steps = 100
rollouts_per_prompt = 4
max_response_length = 512

[resources]
gpu_type = "A100-80GB"
gpu_count = 1

[checkpoint]
destination = "veri"    # or { type = "s3", uri = "s3://..." }

In Phase-1, [dataset].id and [reward].id must reference already-registered IDs. Upload first with veri datasets upload / veri rewards upload, then paste the returned IDs into the config.

`configs/deploy.toml`

kind = "deploy"

[deployment]
name = "my-endpoint"
model = "job_abc"           # training job id, or HF repo
source = "training_job"     # training_job | huggingface

[resources]
gpu_type = "A100-80GB"
gpu_count = 1

[autoscaling]
min_replicas = 1
max_replicas = 1

`configs/eval.toml`

kind = "eval"

[eval]
name = "first-eval"
dataset_id = "ds_test"

[[scorers]]
type = "exact-match"
weight = 1.0

[run]
model = "dep_abc"           # deployment id, training job id, or HF repo
num_samples = 100

`--dry-run` before submitting

On veri jobs create (and veri run with kind="train"), --dry-run validates the config without submitting: It prints the resolved GPU config plus a cost estimate (illustrative values):

$ veri run configs/train.toml --dry-run
Dry run — no job submitted.
┏━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━┓
┃ gpu_type  ┃ gpu_count ┃ provider ┃ duration_minutes ┃ rate_usd_per_hour ┃ estimated_cost_usd ┃ markup ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━┩
│ A100-80GB │ 1         │ aws      │ 100.0            │ 1.79              │ 3.88               │ 1.3    │
└───────────┴───────────┴──────────┴──────────────────┴───────────────────┴────────────────────┴────────┘

Add --format json for the same fields as a JSON object (gpu_type, gpu_count, provider, duration_minutes, rate_usd_per_hour, estimated_cost_usd, markup).

Get started

Training

Deployments

Evaluations

CLI

Volumes

veri run — Hero Verb & Project Layout

Scaffold a project: `veri init`

Dispatch rules

Override TOML fields from the CLI

Type coercion

Repeated flags build a list

Conflict detection

Convenience long-flags

Config schemas

`veri.toml` (project root)

`configs/train.toml`

`configs/deploy.toml`

`configs/eval.toml`

`--dry-run` before submitting

What’s next

Datasets

Training

​Scaffold a project: veri init

​Dispatch rules

​Override TOML fields from the CLI

​Type coercion

​Repeated flags build a list

​Conflict detection

​Convenience long-flags

​Config schemas

​veri.toml (project root)

​configs/train.toml

​configs/deploy.toml

​configs/eval.toml

​--dry-run before submitting

​What’s next

Datasets

Training

Scaffold a project: `veri init`

Dispatch rules

Override TOML fields from the CLI

Type coercion

Repeated flags build a list

Conflict detection

Convenience long-flags

Config schemas

`veri.toml` (project root)

`configs/train.toml`

`configs/deploy.toml`

`configs/eval.toml`

`--dry-run` before submitting

What’s next