🧠 CarsonTalk v0.1
A conversational chatbot fine-tuned on real Discord server messages to emulate casual, chaotic, and humanlike dialogue. Based on distilgpt2.
📦 Model Info
- Base Model: distilgpt2
- Training Data: 2000 Discord messages (anonymized, CSV-based)
- Max Token Length: 128
- Training Time: ~10 minutes on T4
- Epochs: 2
- Batch Size: 8
- Eval Loss: 4.71
- Date: 2025-04-09
📈 Results
| Metric | Value |
|---|
| Eval Loss | 4.71 |
| Avg Msg Len | ~3.36 words |
✨ Sample Outputs
Prompt: "hey what's up"
→ hey what's up with her rlly?
Prompt: "did you see that meme"
→ did you see that meme? i didnt read the full description of how it is in our lives?
Prompt: "i'm not mad but like"
→ i'm not mad but like… you want to ur be in the match...
🧪 Intended Use
- Fun chatbot experiments
- Data-efficient casual dialogue modeling
- Benchmarking micro-finetunes on chaotic social data
📜 License
MIT
Trained and auto-logged with ❤️ in Google Colab.
