Compatibility Validation: The Final Check

You've found a model, downloaded it, converted it, and imported it. But how do you know if it's "broken"? AI models can fail in subtle ways—maybe it's not using the GPU, or maybe the "Stop tokens" are wrong, causing it to ramble forever.

Here is how to validate your new creation.

1. The "First Run" Inspection

Start the model and look for the initial output.

• Is it too fast? (Might be hallucinating gibberish).
• Is it too slow? (Might be accidentally running on the CPU).

Run /show info in the CLI. Check the Architecture and Parameter Count. If it says 0, something went wrong in the Modelfile import.

---

2. Tokenization Validation

The most common error in manual imports is a broken tokenizer. Ask the model: "Tell me your name and what year it is." If the response contains strange symbols (``, [UNK], ###), the tokenizer isn't mapped correctly. You might need to add a TEMPLATE line to your Modelfile (see Module 5).

---

3. The "Infinite Loop" Test

Some models forget to say "End of Turn." They will answer your question and then start talking as "The User" or generate random Python code for 10 minutes. If this happens, you need to find the Stop Token on the Hugging Face page (e.g., <|end_of_text|>) and add it to your Modelfile: PARAMETER stop "<|end_of_text|>"

---

4. Reading the Logs

If Ollama crashes, the answer is in the logs.

• macOS: ~/.ollama/logs/server.log
• Linux: journalctl -u ollama --no-pager
• Windows: Search for "Ollama Logs" in your user folder.

Look for lines containing error or failed to load model. Usually, this indicates you don't have enough VRAM for the specific quantization you chose.

---

5. Summary Checklist

1. GPU Check: Does ollama ps show the model is on the GPU?
2. Format Check: Does output contain weird symbols (``)?
3. Stop Check: Does the model stop generating after it answers?
4. Context Check: Can it remember a 5-turn conversation without crashing?

---

Key Takeaways

• Validation is the process of catching subtle errors in manual imports.
• Tokenization errors are indicated by weird characters or broken formatting.
• Stop tokens prevent the model from rambling in an infinite loop.
• The server log is your best friend when things don't go as planned.