1️⃣ LoRA Fine-tuned Model Testing

Today I tested the LoRA fine-tuned model, but the process did not go smoothly.

Because only partial base model parameters were configured in the config file, and inference-related settings were not fully specified, the result was:

• Inference lasted approximately 5 hours
• High resource consumption
• Unsatisfactory final performance

The main issue stemmed from the evaluation side not imposing limits on timeout and maximum output length.

---

2️⃣ Baseline Model Re-evaluation

I then re-tested a baseline model with a known pass rate and compared different temperature settings:

• temperature = 0.3
• temperature = 0.4

Results:

• At 0.3, the pass rate was approximately 10%
• At 0.4, performance was less stable compared to 0.3

Therefore, subsequent evaluations will adopt temperature = 0.3 as the experimental standard.

---

3️⃣ Reflection

There are still concerns about whether the experiment will eventually produce meaningful results.

However, today’s progress includes:

• Identifying and isolating configuration-related issues
• Determining a more appropriate temperature through comparison
• Keeping the evaluation pipeline logically consistent

Even without breakthrough results, the experiment continues in a structured and disciplined manner.