ReddRadar
Agent skill
training-code-verification
Checklist for editing training code. Trigger when: (1) modifying training notebook, (2) changing GPUEnvConfig, (3) changing reward functions, (4) ANY training-related code changes.
Stars
163
Forks
31
Install this agent skill to your Project
npx add-skill https://github.com/majiayu000/claude-skill-registry/tree/main/skills/development/training-code-verification-smith6jt-cop-skills-registry
SKILL.md
Training Code Verification Checklist
MANDATORY Before Any Training Code Change
DO NOT make assumptions. VERIFY everything.
1. File Paths - CHECK ACTUAL FILES
bash
# List actual files in the repo
ls -la *.txt # API key files
ls -la notebooks/ # Notebook files
ls -la alpaca_trading/gpu/ # Training modules
Common mistakes:
- Assuming
API_key.txtexists (actual:API_key_500Paper.txt,API_key_100kPaper.txt) - Assuming Google Drive paths in Colab (actual:
/content/Alpaca_trading/after unzip) - Assuming relative paths work (use absolute paths)
2. Configuration Values - READ BEFORE CHANGING
python
# ALWAYS read current values first
from alpaca_trading.gpu.vectorized_env import GPUEnvConfig
config = GPUEnvConfig()
# Print ALL values before changing anything
print(f"direction_weight: {config.direction_weight}")
print(f"magnitude_weight: {config.magnitude_weight}")
print(f"pnl_weight: {config.pnl_weight}")
print(f"reward_scale: {config.reward_scale}")
# ... etc
Never assume a value. Read it.
3. Version Strings - VERIFY CONSISTENCY
Check ALL places where version is mentioned:
bash
grep -r "v2\." notebooks/training.ipynb | head -20
All version strings must match. Don't leave old versions.
4. Notebook Cells - CHECK CELL NUMBERS
Notebook cells can shift. Don't assume "cell 15" is always the same cell.
python
# Find cell by content, not number
import json
nb = json.load(open('notebooks/training.ipynb'))
for i, cell in enumerate(nb['cells']):
source = ''.join(cell.get('source', []))
if 'API_KEYS_FILE' in source:
print(f"Cell {i}: API_KEYS_FILE config")
5. Weight Sums - MUST EQUAL 1.0
python
weight_sum = (
config.direction_weight +
config.magnitude_weight +
config.pnl_weight +
config.stop_tp_weight +
config.exploration_weight +
config.slippage_weight +
config.drawdown_penalty_weight
)
assert abs(weight_sum - 1.0) < 0.001, f"Weight sum is {weight_sum}, must be 1.0"
6. Tests - RUN BEFORE COMMITTING
bash
# Run relevant tests
python -m pytest tests/test_ppo_trainer_native.py -v
python -m pytest tests/test_gpu_training.py -v
python -m pytest tests/test_data_fetcher.py -v
7. Colab Workflow - UNDERSTAND THE ENVIRONMENT
| Location | Windows Local | Colab After Unzip |
|---|---|---|
| Repo root | C:\users\smith\Alpaca_trading\ |
/content/Alpaca_trading/ |
| API keys | API_key_500Paper.txt |
/content/Alpaca_trading/API_key_500Paper.txt |
| Notebooks | notebooks\training.ipynb |
/content/Alpaca_trading/notebooks/training.ipynb |
| Drive | N/A | /content/drive/MyDrive/ (separate from repo) |
Failed Attempts Log
| Date | Mistake | Impact | Lesson |
|---|---|---|---|
| 2024-12-28 | Assumed API_key.txt exists |
yfinance fallback used | Check actual filenames |
| 2024-12-28 | Suggested /content/drive/MyDrive/ path |
File not found | Unzip goes to /content/, not Drive |
| 2024-12-28 | Silent exception in key loading | No error shown | Add logging, don't swallow exceptions |
Before Committing Checklist
- Read current values before changing
- Verify file paths exist
- Check all version strings match
- Verify weight sums equal 1.0
- Run relevant tests
- Test in actual environment (Colab if that's target)
- Check logs show expected behavior
Key Principle
Don't assume. Verify.
Every assumption is a potential bug. Read the actual state. Check the actual files. Run the actual tests.
Didn't find tool you were looking for?