DeepSeek API Setup for Continuous Learning GRPO¶

✅ Current Status¶

✅ API Key Saved: .env file created with your DeepSeek API key
✅ Gitignored: .env is ignored by git (won't be committed)
✅ API Key Valid: Key connects successfully to DeepSeek API
⚠️ Balance Required: Account needs credits to run training

💰 Add Credits to DeepSeek Account¶

Your API Key: sk-82accfbadb484ea7ad986510f88d27f5

Steps to Add Credits:

Go to DeepSeek Platform: https://platform.deepseek.com/usage
Log in with your account
Click "Billing" or "Add Credits"
Add credits (minimum $5 recommended)

Cost Estimates: - Quick test (20 samples): ~$0.10** - Full training (100 samples, 3 epochs): **~$0.37 - Pricing: $0.14/1M input tokens, $0.28/1M output tokens

🧪 Test the Setup¶

Once you've added credits, test the API:

cd /Users/z/work/zoo/gym

# Load API key from .env
source .env

# Run API verification test
python test_deepseek_api.py

Expected output:

✓ Client initialized
✓ Response: 8
✓ Summary generated (450 chars)
✓ Extracted 2 operations
✓ Consolidated to 1 final operations
✓ ALL TESTS PASSED - DeepSeek API is working correctly!

🚀 Run Training¶

After credits are added and the test passes:

Quick Test (20 samples, ~$0.10)¶

cd /Users/z/work/zoo/gym
source .env  # Load API key
bash scripts/train_zen_eco_4b_grpo.sh alpaca_en_demo 1 5 20

Full Training (100 samples, ~$0.37)¶

cd /Users/z/work/zoo/gym
source .env  # Load API key
bash scripts/train_zen_eco_4b_grpo.sh alpaca_en 3 5 100

🔒 Security Notes¶

API Key Protection: - ✅ API key is in .env (gitignored) - ✅ Never commit .env to git - ✅ Never hardcode API keys in scripts - ✅ Always use source .env to load keys

To verify gitignore:

git status .env
# Should show: ".env is gitignored"

To check if .env would be committed:

git check-ignore .env
# Should output: ".env"

🛠️ Troubleshooting¶

Error: "Insufficient Balance" (402)¶

Problem: DeepSeek account has no credits
Solution: Add credits at https://platform.deepseek.com/usage

Error: "No LLM API key found"¶

Problem: API key not loaded from .env
Solution: Run source .env before training

Error: "Invalid API key" (401)¶

Problem: Wrong API key or expired
Solution: Generate new key at https://platform.deepseek.com/api_keys

To check if .env is loaded:¶

echo $DEEPSEEK_API_KEY
# Should show: sk-82accfbadb484ea7ad986510f88d27f5

📊 Expected Performance¶

Based on Tencent Paper Results:

Metric	Value
Performance Gain	+1-3% on math tasks
Training Cost	$0.37 for 100 samples
Training Time	~10 minutes
Experiences Generated	50-100 after 3 epochs

📝 Next Steps¶

Add Credits: https://platform.deepseek.com/usage (minimum $5)
Test API: source .env && python test_deepseek_api.py
Run Training: bash scripts/train_zen_eco_4b_grpo.sh alpaca_en_demo 1 5 20
Analyze Results: Check ./output/zen-eco-4b-grpo/experiences.json

Last Updated: October 28, 2025
API Key: sk-82accfbadb484ea7ad986510f88d27f5
Status: Ready (needs credits)