Hosted Training

Train models on Prime Intellect’s managed infrastructure

Hosted Training, available within the Prime Intellect Lab platform, enables you to automatically train models via prime-rl without needing to manage your own infrastructure.

Hosted Training supports LoRA for RL training and can be used with any environment built with Verifiers.

Features

• Zero infrastructure management - No need to provision GPUs or manage servers
• Automatic scaling - Training infrastructure scales based on your config
• LoRA training - Efficient parameter-efficient fine-tuning
• Any Verifiers environment - Train on Hub environments or your own
• Weights & Biases integration - Automatic logging and experiment tracking

Getting Started

Hosted Training is currently in Private Beta. For access, please fill out this form.

1. Set up your workspace

   Download example configuration files:

   prime lab setup
   

   This creates:

   configs/
   ├── endpoints.toml      # API endpoint configuration
   ├── rl/                 # Hosted Training configs
   │   ├── alphabet-sort.toml
   │   ├── gsm8k.toml
   │   ├── math-python.toml
   │   ├── reverse-text.toml
   │   ├── wiki-search.toml
   │   └── wordle.toml
   ├── eval/               # Evaluation configs
   └── gepa/               # Prompt optimization configs
   
2. Configure your training

   Edit one of the example configs or create your own. Example for alphabet-sort:

   model = "Qwen/Qwen3-30B-A3B-Instruct-2507"
   max_steps = 500
   batch_size = 256
   rollouts_per_example = 8
   
   [sampling]
   max_tokens = 512
   
   [[env]]
   id = "primeintellect/alphabet-sort"
   args = { min_turns = 3, max_turns = 5, power_per_turn = false }
   
   [wandb]
   project = "alphabet-sort"
   name = "qwen3-30b-i-alphabet-sort"
   
3. Submit your training job

   Submit to Hosted Training via the Prime CLI:

   prime train submit configs/rl/alphabet-sort.toml
   

   Or use the web interface at app.primeintellect.ai/dashboard/training

4. Monitor your training

   View training progress:

   • In the Prime Intellect dashboard
   • In Weights & Biases (if configured)
   • Via the CLI: prime train status <job-id>

Supported Models

We currently support the following models for Hosted Training:

• Qwen/Qwen3-4B-Instruct-2507
• Qwen/Qwen3-4B-Thinking-2507
• Qwen/Qwen3-30B-Instruct-2507
• Qwen/Qwen3-30B-Thinking-2507
• Qwen/Qwen3-235B-Instruct-2507
• Qwen/Qwen3-235B-Thinking-2507
• PrimeIntellect/INTELLECT-3

Additional models can be supported upon request. Contact support if you need a specific model.

Configuration Reference

Basic Configuration

model = "Qwen/Qwen3-4B-Instruct-2507"
max_steps = 500
batch_size = 256
rollouts_per_example = 8
learning_rate = 1e-5

Environment Configuration

Train on environments from the Environments Hub:

[[env]]
id = "primeintellect/math-python"
args = { max_turns = 10, difficulty = "hard" }

Or train on your own local environment:

[[env]]
id = "my-custom-env"  # from ./environments/my_custom_env
args = { num_examples = 1000 }

Multiple environments:

[[env]]
id = "primeintellect/math-python"
weight = 0.6

[[env]]
id = "primeintellect/gsm8k"
weight = 0.4

Sampling Configuration

[sampling]
max_tokens = 512
temperature = 0.7
top_p = 0.9
stop = ["<|endoftext|>", "</s>"]

LoRA Configuration

LoRA is enabled by default for Hosted Training:

[lora]
enabled = true
r = 64
alpha = 16
dropout = 0.05
target_modules = ["q_proj", "v_proj", "k_proj", "o_proj"]

Weights & Biases Integration

[wandb]
project = "my-project"
name = "my-training-run"
entity = "my-team"  # optional

Set your W&B API key in the Prime Intellect dashboard under Settings > Environment Variables.

Environment Variables

If your environment requires API keys or secrets, configure them via:

1. Dashboard: Settings > Environment Variables
2. Config file:

env_file = ["secrets.env"]

Then create secrets.env:

OPENAI_API_KEY=sk-...
BROWSERBASE_API_KEY=...

Environment variables set in the dashboard take precedence over those in env_file.

Training Examples

GSM8K Math Training

model = "Qwen/Qwen3-4B-Instruct-2507"
max_steps = 1000
batch_size = 256
rollouts_per_example = 8

[sampling]
max_tokens = 1024
temperature = 0.7

[[env]]
id = "primeintellect/gsm8k"
args = { max_turns = 1 }

[wandb]
project = "gsm8k-training"

Multi-Turn Wiki Search

model = "Qwen/Qwen3-30B-Instruct-2507"
max_steps = 500
batch_size = 128
rollouts_per_example = 8

[sampling]
max_tokens = 2048

[[env]]
id = "primeintellect/wiki-search"
args = { max_turns = 10, num_questions = 5 }

[wandb]
project = "wiki-search"
name = "qwen3-30b-wiki"

Multi-Environment Training

model = "Qwen/Qwen3-4B-Instruct-2507"
max_steps = 2000
batch_size = 256

[[env]]
id = "primeintellect/math-python"
weight = 0.4

[[env]]
id = "primeintellect/gsm8k"
weight = 0.3

[[env]]
id = "primeintellect/wiki-search"
weight = 0.3

[wandb]
project = "multi-task-training"

Downloading Checkpoints

After training completes, download your trained model:

prime train download <job-id> --output ./checkpoints/my-model

This downloads the final checkpoint and LoRA adapter (if applicable).

Best Practices

Before submitting a training job, validate your environment locally:

prime eval run my-env -m openai/gpt-4.1-mini -n 10

Ensure baseline reward is between 5% and 80%.

Hyperparameter Guidelines

For faster training:

• Use smaller models (4B-30B)
• Increase learning rate (1e-5 to 1e-4)
• Decrease rollouts_per_example (4-8)

For more stable training:

• Use larger models (30B+)
• Increase rollouts_per_example (16-32)
• Increase batch_size (512+)

Cost Optimization

• Use LoRA instead of full finetuning
• Start with smaller models and scale up if needed
• Use max_steps to limit training duration
• Monitor W&B to stop training when performance plateaus

Troubleshooting

Training Not Starting

• Check that your config is valid TOML
• Ensure your environment is published to the Environments Hub (if using a Hub environment)
• Verify all required API keys are set

Training Failed

• Check job logs: prime train logs <job-id>
• Common issues:
  • Missing environment dependencies
  • Invalid environment arguments
  • Missing API keys for environment

Poor Training Performance

• Task may be too hard for the model (baseline reward < 5%)
• Task may be too easy (baseline reward > 80%)
• Learning rate may be too high (causing instability)
• Try enabling online difficulty filtering in advanced settings

Support

For help with Hosted Training:

• Email: [e-mail]
• Discord: discord.gg/primeintellect
• Documentation: docs.primeintellect.ai