Training ↗

Documentation Index#

Fetch the complete documentation index at: https://docs.crewai.com/llms.txt Use this file to discover all available pages before exploring further.

Learn how to train your CrewAI agents by giving them feedback early on and get consistent results.

Overview#

The training feature in CrewAI allows you to train your AI agents using the command-line interface (CLI). By running the command crewai train -n <n_iterations>, you can specify the number of iterations for the training process.

During training, CrewAI utilizes techniques to optimize the performance of your agents along with human feedback. This helps the agents improve their understanding, decision-making, and problem-solving abilities.

Training Your Crew Using the CLI#

To use the training feature, follow these steps:

1. Open your terminal or command prompt.
2. Navigate to the directory where your CrewAI project is located.
3. Run the following command:

crewai train -n <n_iterations> -f <filename.pkl>

Replace <n_iterations> with the desired number of training iterations and <filename> with the appropriate filename ending with .pkl.

If you omit -f, the output defaults to trained_agents_data.pkl in the current working directory. You can pass an absolute path to control where the file is written.

Training your Crew programmatically#

To train your crew programmatically, use the following steps:

1. Define the number of iterations for training.
2. Specify the input parameters for the training process.
3. Execute the training command within a try-except block to handle potential errors.

n_iterations = 2
inputs = {"topic": "CrewAI Training"}
filename = "your_model.pkl"

try:
    YourCrewName_Crew().crew().train(
      n_iterations=n_iterations,
      inputs=inputs,
      filename=filename
    )

except Exception as e:
    raise Exception(f"An error occurred while training the crew: {e}")

How trained data is used by agents#

CrewAI uses the training artifacts in two ways: during training to incorporate your human feedback, and after training to guide agents with consolidated suggestions.

Training data flow#

flowchart TD
  A["Start training<br/>CLI: crewai train -n -f<br/>or Python: crew.train(...)"] --> B["Setup training mode<br/>- task.human_input = true<br/>- disable delegation<br/>- init training_data.pkl + trained file"]

  subgraph "Iterations"
    direction LR
    C["Iteration i<br/>initial_output"] --> D["User human_feedback"]
    D --> E["improved_output"]
    E --> F["Append to training_data.pkl<br/>by agent_id and iteration"]
  end

  B --> C
  F --> G{"More iterations?"}
  G -- "Yes" --> C
  G -- "No" --> H["Evaluate per agent<br/>aggregate iterations"]

  H --> I["Consolidate<br/>suggestions[] + quality + final_summary"]
  I --> J["Save by agent role to trained file<br/>(default: trained_agents_data.pkl)"]

  J --> K["Normal (non-training) runs"]
  K --> L["Auto-load suggestions<br/>from trained_agents_data.pkl"]
  L --> M["Append to prompt<br/>for consistent improvements"]

During training runs#

• On each iteration, the system records for every agent:
  • initial_output: the agent’s first answer
  • human_feedback: your inline feedback when prompted
  • improved_output: the agent’s follow-up answer after feedback
• This data is stored in a working file named training_data.pkl keyed by the agent’s internal ID and iteration.
• While training is active, the agent automatically appends your prior human feedback to its prompt to enforce those instructions on subsequent attempts within the training session. Training is interactive: tasks set human_input = true, so running in a non-interactive environment will block on user input.

After training completes#

• When train(...) finishes, CrewAI evaluates the collected training data per agent and produces a consolidated result containing:
  • suggestions: clear, actionable instructions distilled from your feedback and the difference between initial/improved outputs
  • quality: a 0–10 score capturing improvement
  • final_summary: a step-by-step set of action items for future tasks
• These consolidated results are saved to the filename you pass to train(...) (default via CLI is trained_agents_data.pkl). Entries are keyed by the agent’s role so they can be applied across sessions.
• During normal (non-training) execution, each agent automatically loads its consolidated suggestions and appends them to the task prompt as mandatory instructions. This gives you consistent improvements without changing your agent definitions.

File summary#

• training_data.pkl (ephemeral, per-session):
  • Structure: agent_id -> { iteration_number: { initial_output, human_feedback, improved_output } }
  • Purpose: capture raw data and human feedback during training
  • Location: saved in the current working directory (CWD)
• trained_agents_data.pkl (or your custom filename):
  • Structure: agent_role -> { suggestions: string[], quality: number, final_summary: string }
  • Purpose: persist consolidated guidance for future runs
  • Location: written to the CWD by default; use -f to set a custom (including absolute) path

Small Language Model Considerations#

When using smaller language models (≤7B parameters) for training data evaluation, be aware that they may face challenges with generating structured outputs and following complex instructions.

Limitations of Small Models in Training Evaluation#

<span class=“card-start” data-card-raw=“title=“JSON Output Accuracy” icon=“triangle-exclamation”"> Smaller models often struggle with producing valid JSON responses needed for structured training evaluations, leading to parsing errors and incomplete data.

<span class=“card-start” data-card-raw=“title=“Evaluation Quality” icon=“chart-line”"> Models under 7B parameters may provide less nuanced evaluations with limited reasoning depth compared to larger models.

<span class=“card-start” data-card-raw=“title=“Instruction Following” icon=“list-check”"> Complex training evaluation criteria may not be fully followed or considered by smaller models.

<span class=“card-start” data-card-raw=“title=“Consistency” icon=“rotate”"> Evaluations across multiple training iterations may lack consistency with smaller models.

Recommendations for Training#

For optimal training quality and reliable evaluations, we strongly recommend using models with at least 7B parameters or larger:

    from crewai import Agent, Crew, Task, LLM

    # Recommended minimum for training evaluation
    llm = LLM(model="mistral/open-mistral-7b")

    # Better options for reliable training evaluation
    llm = LLM(model="anthropic/claude-3-sonnet-20240229-v1:0")
    llm = LLM(model="gpt-4o")

    # Use this LLM with your agents
    agent = Agent(
        role="Training Evaluator",
        goal="Provide accurate training feedback",
        llm=llm
    )
    ```

<span class="callout-start" data-callout-type="tip"></span>
  More powerful models provide higher quality feedback with better reasoning, leading to more effective training iterations.
<span class="callout-end"></span>
  <span class="tab-end"></span>

  <span class="tab-start" data-tab-title="Small Model Usage"></span>
If you must use smaller models for training evaluation, be aware of these constraints:

```python
    # Using a smaller model (expect some limitations)
    llm = LLM(model="huggingface/microsoft/Phi-3-mini-4k-instruct")
    ```

<span class="callout-start" data-callout-type="warning"></span>
  While CrewAI includes optimizations for small models, expect less reliable and less nuanced evaluation results that may require more human intervention during training.
<span class="callout-end"></span>
  <span class="tab-end"></span>
<span class="tab-group-end"></span>

### Key Points to Note

* **Positive Integer Requirement:** Ensure that the number of iterations (`n_iterations`) is a positive integer. The code will raise a `ValueError` if this condition is not met.
* **Filename Requirement:** Ensure that the filename ends with `.pkl`. The code will raise a `ValueError` if this condition is not met.
* **Error Handling:** The code handles subprocess errors and unexpected exceptions, providing error messages to the user.
* Trained guidance is applied at prompt time; it does not modify your Python/YAML agent configuration.
* Agents automatically load trained suggestions from a file named `trained_agents_data.pkl` located in the current working directory. If you trained to a different filename, either rename it to `trained_agents_data.pkl` before running, or adjust the loader in code.
* You can change the output filename when calling `crewai train` with `-f/--filename`. Absolute paths are supported if you want to save outside the CWD.

It is important to note that the training process may take some time, depending on the complexity of your agents and will also require your feedback on each iteration.

Once the training is complete, your agents will be equipped with enhanced capabilities and knowledge, ready to tackle complex tasks and provide more consistent and valuable insights.

Remember to regularly update and retrain your agents to ensure they stay up-to-date with the latest information and advancements in the field.