webapp/COLLABORATION.md

🤝 Team Collaboration Guide

Quick Reference for Updating the Deployed App

Space URL: https://huggingface.co/spaces/dianacasti/personality-chatbot

---

I just trained a better adapter. How do I update it?

Option A: Direct Upload (No Git Required) ⭐ EASIEST

1. Ask dianacasti to add you as a collaborator on the Space
2. Go to: https://huggingface.co/spaces/dianacasti/personality-chatbot
3. Click "Files" tab at the top
4. Find the old adapter folder (e.g., nerd-lora-adapter)
5. Click the trash icon to delete it
6. Click "Add file" → "Upload files"
7. Drag and drop your entire new adapter folder OR upload files individually:
   • adapter_model.safetensors (the weights)
   • adapter_config.json
   • tokenizer.json, tokenizer_config.json, vocab.json, merges.txt
   • special_tokens_map.json, added_tokens.json
   • chat_template.jinja
8. Write a commit message: "Updated nerd adapter with 50 training examples"
9. Click "Commit changes to main"
10. Wait 2-5 minutes for rebuild
11. Test at: https://dianacasti-personality-chatbot.hf.space

Done! ✅

---

Option B: Using Git (For Advanced Users)

# One-time setup
git clone https://huggingface.co/spaces/dianacasti/personality-chatbot
cd personality-chatbot

# Every time you update
# 1. Copy your new adapter folder to replace the old one
cp -r C:\path\to\your\nerd-lora-adapter .\nerd-lora-adapter

# 2. Commit and push
git add .
git commit -m "Updated nerd adapter - trained with 50 examples, loss 0.018"
git push

# 3. Wait for rebuild and test

---

Common Tasks

✅ Replace an existing adapter (Nerd, Yoda, Pirate)

• Use Option A above (direct upload is easiest)

✅ Test adapters locally before uploading

cd C:\Users\dcast\LLM\Project3\llm-project

# Replace the adapter folder with your new one
# Then run locally:
python app.py

# Test at http://localhost:7860
# If good, upload to Hugging Face using Option A

✅ Check current training loss/quality

• Nerd: loss 0.018 (50 examples × 10)
• Yoda: loss 0.0165 (50 examples × 10)
• Brainrot: loss 0.21 (large dataset)
• Pirate: original (5 examples)

Lower loss = better memorization, but might overfit if too low

✅ Add a brand new personality

1. Train your new adapter (e.g., cowboy-lora-adapter)
2. Upload the adapter folder using Option A
3. Edit app.py (Files tab → click app.py → edit icon):

   ADAPTERS = {
       "🧠 Brainrot": "qwen-brainrot-lora-stage1-final",
       "🏴‍☠️ Pirate": "pirate-lora-adapter",
       "🧙 Yoda": "yoda-lora-adapter",
       "🤓 Nerd": "nerd-lora-adapter",
       "🤠 Cowboy": "cowboy-lora-adapter",  # ADD THIS LINE
   }
   

4. Scroll down to personality descriptions and add:

   - **🤠 Cowboy**: Yeehaw, talks like a Wild West cowboy!
   

5. Commit changes

---

Troubleshooting

"I don't see the Files tab"

→ You need to be added as a collaborator. Ask dianacasti.

"Space is stuck building"

→ Check the "Logs" tab for errors. Common issues:

• Missing files in adapter folder
• File names don't match (should be adapter_model.safetensors, not model.safetensors)

"Adapter not loading"

→ Make sure adapter folder contains ALL files:

• adapter_model.safetensors ✅
• adapter_config.json ✅
• tokenizer files (5-6 files) ✅
• chat_template.jinja ✅

"How do I test without breaking the live app?"

→ Clone locally and test with python app.py before uploading

---

Training Tips for Better Adapters

Current approach: 5 examples × 50 repetitions = 250 training samples (overfits easily)

Better approach: 50 unique examples × 3 epochs

• More diverse conversations
• Less overfitting
• Better generalization

Example training data structure:

train_data = [
    {"messages": [
        {"role": "user", "content": "Hello"},
        {"role": "assistant", "content": "Ahoy matey!"}
    ]},
    # ... 49 more unique examples
] * 3  # Repeat 3 times instead of 50

---

Contact

• Space Owner: dianacasti
• Project Lead: Theworst1
• Contributors: Bryn Ramirez, diana

For issues or questions, post in the team Discord!