Training Log

Failure → fix note

Failure pattern: early generations sounded repetitive and overly rehearsed.

Likely cause: the tiny corpus had repeated phrasing and the initial tuning plan was too aggressive for such a narrow dataset. In the module’s language, this looked like the beginning of narrow specialization pressure with catastrophic forgetting risk.

Fix: keep exact-duplicate removal in the data pipeline, use a conservative learning rate (5e-5), preserve a held-out split, and compare base vs tuned generations at the same temperatures.

Evidence to watch:

• held-out perplexity improves instead of only train loss
• tuned generations pick up domain vocabulary without losing basic coherence
• base vs tuned comparison stays fair because decoding settings match

Notes to capture after your run

• Base perplexity:
• Tuned perplexity:
• Perplexity delta:
• Best sample improvement:
• Worst regression:
• Next change to try: