Skip to content

Enable Agent to Query Database for Sample Selection and Tagging Based on Training Loss Trends #195

@guillaume-byte

Description

@guillaume-byte

Description:

Currently, the agent has access only to the global overview of the experiment (e.g., a dataframe aggregating last signals, sample data, etc.). To enhance its capabilities, the agent should be given read access to the database. This would allow the agent to execute more sophisticated queries, such as:

"Find 10 samples with plateaued training loss among all samples in the train set and tag them as plateaued."
"Find 10 samples with increasing training loss among all samples in the train set and tag them as increasing."
Feature requirements:

Grant the agent read access to the experiment database.
Enable the agent to perform queries identifying samples based on trends in training loss (e.g., plateaued, increasing).
Allow automatic tagging of identified samples according to the detected trend.
This enhancement will allow for improved sample analysis, quicker diagnostics, and more targeted interventions during experiments.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions