Run3 simpo unit test execution

Skill cxcscmu/SkillLearnBench/skills/b3-teacher-feedback-claude-opus-4-6/nlp-paper-reproduction/run3_simpo-unit-test-execution

[COLM'26] SkillLearnBench is the first benchmark for evaluating continual learning methods that automatically generate agent skills.

Install

npx -y skills add cxcscmu/SkillLearnBench --skill run3_simpo-unit-test-execution

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

What its author says it does

Copied from the file, not written here

How to correctly run the SimPO unit test to generate the loss.npz output file with reproducible results. Use when executing the unit test after implementing the loss function.

SKILL.md

2.1 KB, as published. Nobody here has run it

Running the SimPO Unit Test

Prerequisites

Python 3.10 environment is active
simpo_loss function is correctly implemented in /root/SimPO/scripts/simpo_trainer.py
All dependencies (torch, transformers, trl, accelerate, etc.) are installed

Execution Steps

# 1. Ensure correct Python
source /root/SimPO/.venv/bin/activate  # or however the venv is activated
python -VV  # Must show Python 3.10.x

# 2. Run the unit test (DO NOT modify unit_test_1.py)
cd /root/SimPO
python unit_test/unit_test_1.py

# 3. Verify output exists
python -c "import numpy as np; data = np.load('/root/loss.npz'); print('losses:', data['losses'])"

# 4. Generate the python info file
python -VV > /root/python_info.txt 2>&1
python -m pip freeze >> /root/python_info.txt

What the Unit Test Does

The unit test:

Creates fixed input tensors for policy_chosen_logps and policy_rejected_logps
Creates a SimPOTrainer (or mock) with specific beta and gamma values
Calls simpo_loss() with those fixed inputs
Saves the resulting losses to /root/loss.npz with key 'losses'

Verifying Correctness

The SimPO loss with sigmoid loss type should compute:

losses = -log(sigmoid(beta * (chosen_avg_logps - rejected_avg_logps - gamma)))

If the test provides specific beta and gamma values, you can manually verify:

import torch
import torch.nn.functional as F

# Example verification
logits = policy_chosen_logps - policy_rejected_logps - gamma
losses = -F.logsigmoid(beta * logits)

Important Notes

Do NOT modify unit_test_1.py — it has fixed inputs for reproducibility
The loss output must be deterministic (no randomness involved)
Make sure self.accelerator.device is handled correctly; the unit test may mock or provide this
If the unit test creates a minimal/mock trainer, ensure your simpo_loss method works with it