Another round of "chainsaw°hard word coding°hard word" with Claude°hard word. 🔥 the generators°hard word now use generators°hard word! 💡 Previously, Claude's°hard word approach°hard word was to just make one question without history. But we do want history. Use all the canned°hard word questions from text°hard word files°hard word (once), then°hard word generate files°hard word from local°hard word data (without repeating°hard word), then°hard word ask the LLM°hard word.
It mostly works. The LLM°hard word generation°hard word is slightly flaky°hard word. It gave°hard word both "sad" and "melancholy°hard word" as°hard word candidate°hard word opposites for "happy". It is including°hard word the pinyin°hard word for Chinese°hard word translations°hard word some of the time.
But at least it is (mostly) using°hard word the correct interfaces°hard word now. 💡 there are still some useless°hard word parameters°hard word. The "tags" for questions are meaningless°hard word. The difficulty°hard word is arbitrary°hard word. The evaluation°hard word criteria°hard word are often silly°hard word.
----
The various°hard word "general knowledge" benchmarks°hard word should be quick to write 🔥 tomorrow. Right now I see six categories°hard word for the initial°hard word tests:
- History
- Geography
- Chemistry
- Biology
- Sports 🔥 American athletes°hard word, mostly
- Music 💡 20th°hard word century English-language°hard word music, mostly
These will be "easy" questions. Or, at least, multiple-choice°hard word.