Subjective Evaluations of Trillion's 7B Preview
Answers well and translates well
Croatian language is disaster, German looks good so far.
I see it quoting some sentences even though they were not quoted, which is for me unusual. I am using many models. Prompt is:
Correct the following text strictly for grammar, spelling, punctuation, and clarity without altering its original meaning. Do not provide any notes, explanations, or comments—only return the corrected text. Do not quote if not necessary.
Text
It can't handle sarcasm
Prompt
OUTPUT INSTRUCTIONS
Do not output notes, comments, explanations, descriptions.
Do not output warnings or notes
Convert the style of the following text to sarcastic and provide single emoji.
But no other model understands it so to output 😊 on input like "It can't handle sarcasm". I will change the prompt.
When translating, often no space after colon, or no dot on the end of the sentence. This behavior I found on SmolVLM from Hugging Face as well
9.9 is smaller than 9.11.
“which number is smaller, 9.9 or 9.11?”
This question is answered incorrectly, and Phi-3.5-mini is good, but most of models get it wrong.
Great for horoscope!
It gives such good answers, it seems it is great for horoscope. It doesn't even mention that horoscope is just for entertainment and not objectively real.
https://gnu.support/files/tmp/clipboard-2025-04-02-18-43-02.html
More answers from superstition category:
https://gnu.support/files/tmp/clipboard-2025-04-02-18-43-40.html
https://gnu.support/files/tmp/clipboard-2025-04-02-18-44-23.html
It could not understand my wife sleeps, that I could ask her for the phone:
https://gnu.support/files/tmp/clipboard-2025-04-02-18-47-19.html