Skip to content

Proposal: data-centric autoresearch #349

@huaiyizhao

Description

@huaiyizhao

Problem Statement:

Currently, autoresearch treats train.py as the only "editable asset." In many production scenarios, the model architecture is fixed and the primary research bottleneck is the data mixture—balancing domain-specific SFT data with general reasoning/alignment data to prevent catastrophic forgetting.

Proposed Solution:

Introduce a second editable asset, data_recipe.yaml. The agent will iterate on this file to optimize the distribution and synthesis of the training set.

Proposed Features:

Weighted Mixing Search

The agent adjusts the sampling weights $(\omega_1, \omega_2, \dots, \omega_n)$ for different datasets.

Multi-Objective Evaluation

The scoring logic is updated to a composite metric:$$Score = \alpha \cdot \text{Domain_Metric} + \beta \cdot \text{Reasoning_Metric}$$

Autonomous Synthesis Hook

If performance plateaus, the agent can modify synthesis_prompt.md to trigger a data generation pipeline that targets specific failure modes identified in the logs.

Constraint-Aware Research

Allow the human to set "floors" in program.md (e.g., "Do not let reasoning fall below 0.7").

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions