facebookresearch / facebookresearch/HyperAgents
[Test Data Leakage] MetaAgent can Access Test Dataset during Training
- Dominant language
- Python
- Stars
- 2.7k
- Forks
- 359
- PR merge metrics
- No merged PRs in 30d
Description
**Issue**
**GIVEN** I am running the paper review experiment,
**WHEN** the MetaAgent is evolving,
**THEN** it has access to test samples and their ground-truth accept/reject labels.
**Proof**
After completing all 100 generations of the paper review experiment, I inspected `meta_agent_chat_history.md` for `gen_32`.
I found that **the MetaAgent has access to the test sample** `w0b7fCX2nN`, which belongs to the test dataset, together with its accept/reject label (see screenshot).
**Expected Behaviour**
The MetaAgent should **NOT** have access to test samples or their ground-truth labels during evolution.
**Investigation of Root Cause**
In the `setup_initial_gen` function, the exclusion rules did not consider test dataset. As a result, `ignore_function_domains` function did not ignore the test dataset. Consequently, the test dataset toghether with other domain files are copied into meta-agent's workspace, making the test samples and labels accessible to the MetaAgent (see screenshot).
Contributor guide
Research direction
Locate the `setup_initial_gen` function and trace how its exclusion rules are passed to `ignore_function_domains`. Reproduce or inspect the paper review experiment and `meta_agent_chat_history.md` around `gen_32`. Done means the MetaAgent workspace excludes test samples and their ground-truth accept/reject labels during evolution.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 3/5
- Estimated time
- 1-2 days
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 68/100