| name | run-mlp-autoresearch |
| description | Autonomously iterate and improve the MLP model in MLP-AutoResearch. |
| argument-hint | Run a series of ML experiments |
| user-invocable | true |
MLP-AutoResearch Automation
When to Use
- When asked to autonomously run experiments and improve the MLP model in
MLP-AutoResearch.
Context
This is an AutoResearch framework scaled down to an MLP on MNIST. The goal is to maximize test_accuracy while staying within the constraints (fixed epochs, dataset).
Procedure
For each iteration (repeat as many times as requested by the user):
- Analyze Data: Read
program.md, runner.py, and train.py to understand the current architecture and allowed changes. View results.tsv if it exists.
- Propose Modification: Think of an improvement (e.g. optimizer, activation function, dropout, hidden size, learning rate schedule, layer normalization).
- Apply Code Change: Modify
train.py using file editing tools to implement the proposed idea. Ensure no infrastructure (like prepare.py) is altered.
- Run Experiment: Execute
python train.py in the terminal and wait for it to complete.
- Evaluate Results: Read the output. Look for
best_test_accuracy and compare it to previous baselines.
- Log Result: Write a summary of the change, test accuracy, parameters, and whether the change was kept or reverted to
results.tsv. If the result is worse, revert the change in train.py.