> ## Documentation Index
> Fetch the complete documentation index at: https://docs.trymonte.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Recipes

> The YAML file behind each training run

A recipe is a versioned YAML file that sets a run's base model, method, and hyperparameters. The run names its training data, step count, and starting point with flags, so one recipe serves many runs.

```yaml triage-sft.yaml theme={null}
model: Qwen/Qwen3-8B
method: sft

lora:
  rank: 64

batch_size: 32
learning_rate: 1.0e-4
max_seq_length: 4096
checkpoint_every: 50
```

| Key | Default | Description |
| - | - | - |
| `model` | required | The base model, as a Hugging Face ID from [Models](https://docs.trymonte.ai/training/models.md) |
| `method` | required | `sft`, `dpo`, `grpo`, or `distillation` |
| `lora` | none: a full fine-tune | Train a LoRA adapter. An empty block takes the defaults below. |
| `lora.rank` | `16` | The adapter rank. It must not exceed the `--max-adapter-rank` of the [base deployment](https://docs.trymonte.ai/serving/deploy.md) that serves it. |
| `lora.alpha` | twice the rank | The scale of the adapter update |
| `batch_size` | required | Rows, pairs, or prompts in each step |
| `learning_rate` | required | The highest learning rate of the schedule, after warmup |
| `max_seq_length` | `4096` | The token limit for one row or rollout |
| `checkpoint_every` | `50` | The step interval between saved checkpoints |

Every run trains with AdamW, a linear warmup over the first 3% of steps, then cosine decay, and weight decay 0.0. A `lora` block applies the adapter to all linear layers, with no dropout. No recipe key changes these settings.

Each method adds its own keys under its name, such as `grpo.group_size`. The method pages list them: [SFT](https://docs.trymonte.ai/training/sft.md), [GRPO](https://docs.trymonte.ai/training/grpo.md), [DPO](https://docs.trymonte.ai/training/dpo.md), and [Distillation](https://docs.trymonte.ai/training/distillation.md).

`--recipe triage-sft` finds `recipes/triage-sft.yaml`, and `--recipe` also takes a path. Each run records the hash of the file, so an edit makes a new version of the recipe. `--model` replaces `model` for one run.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.