Feedback and Training Data

Normal AI requests are not retained by Zed. For Zed-hosted models, provider agreements require zero data retention and prohibit training on your prompts or code context. This page covers the cases where Zed may retain AI data because you explicitly shared it or opted in.

AI features in Zed include:

• Agent Panel
• Edit Prediction
• Inline Assistant
• Git commit generation

For the broader request path and provider data boundaries, see AI Privacy.

Zed-Hosted Model Commitments

Zed-hosted model zero-data-retention and no-training commitments are documented on AI Privacy.

Response Ratings and Feedback

You can rate AI responses or submit feedback to help improve Zed's system prompt, tools, and AI product experience. Each share is opt-in, and sharing once does not grant permission for future collection.

Warning: Rating an AI response sends the conversation thread to Zed. The conversation thread includes your messages, AI responses, and thread metadata. If you do not want the thread persisted by Zed, do not rate the response.

Data Collected from Feedback

For conversation threads you explicitly share through ratings or feedback, Zed may store:

• your messages and AI responses in the conversation thread
• any commentary you include with your rating or feedback
• thread metadata, such as model used, token counts, and timestamps
• metadata about your Zed installation

If you do not rate responses or submit feedback, Zed does not store Customer Data related to your AI feature usage for improvement.

Telemetry related to Zed's AI features is collected separately. This includes metadata such as the AI feature being used and high-level interactions with the feature to understand performance, such as agent response time or edit acceptance and rejection. See Telemetry for details.

Collected feedback data is stored in Snowflake. Zed periodically reviews this data to refine prompts, tools, and product behavior. Stored feedback data is anonymized and stripped of sensitive information such as access tokens, user IDs, and email addresses.

Edit Prediction Training Data

Zed does not collect training data for the Edit Prediction model unless all of these conditions are met:

1. You opt in by toggling Training Data Collection under the Privacy section of the Edit Prediction status bar menu.
2. The project is open source, detected by the presence of a license file. See the license detection logic.
3. The file is not excluded by edit_predictions.disabled_globs.

Edit Prediction setup and provider configuration live on the Edit Prediction page. This page only covers training data collection and retention.

File Exclusions

Certain files are always excluded from Edit Prediction training data collection, regardless of opt-in status:

{
  "edit_predictions": {
    "disabled_globs": [
      "**/.env*",
      "**/*.pem",
      "**/*.key",
      "**/*.cert",
      "**/*.crt",
      "**/.dev.vars",
      "**/secrets.yml"
    ]
  }
}

You can explicitly exclude additional paths or file extensions by adding them to edit_predictions.disabled_globs in your Zed settings file (how to edit):

{
  "edit_predictions": {
    "disabled_globs": ["secret_dir/*", "**/*.log"]
  }
}

Data Collected from Edit Prediction Training

For open source projects where you opted in, Zed may collect:

• code excerpts around your cursor
• recent edit diffs
• the generated prediction
• repository URL and git revision
• buffer outline and diagnostics

Collected data is stored in Snowflake. Zed periodically reviews this data to select training samples for inclusion in the model training dataset. Included data is anonymized and stripped of sensitive information such as access tokens, user IDs, and email addresses.

The training dataset is publicly available at huggingface.co/datasets/zed-industries/zeta.

Training Dataset and Model Output

Zed uses this training dataset to fine-tune Qwen2.5-Coder-7B and makes the resulting model available at huggingface.co/zed-industries/zeta.

Business Controls

On Zed Business, data sharing is off by default and controlled by organization administrators. Administrators can prevent members from submitting agent thread feedback or opting into Edit Prediction training data collection. See Privacy for Business and Admin Controls.

Applicable Terms

See the Zed Terms of Service for more.