ILikeAI / AlwaysReddy

AlwaysReddy is a LLM voice assistant that is always just a hotkey away.
MIT License
524 stars 50 forks source link

More reliable START_SEQ and END_SEQ defaults #64

Closed Jobus0 closed 3 weeks ago

Jobus0 commented 3 weeks ago

Change the default config clipboard START_SEQ and END_SEQ to "[CLIPSTART]" and "[CLIPEND]" for better reliability.

To find this, I ran multiple styles through a benchmark of 20 different questions repeated 200 times for each style.

Model: Llama 3 8B START_SEQ END_SEQ Success rate (%)
-CLIPSTART- -CLIPEND- 82.5% (current default)
CLIPSTART CLIPEND 94%
#CLIPSTART #CLIPEND 92%
[CLIPSTART] [CLIPEND] 96.5% (best)
\<CLIPSTART> \<CLIPEND> 93.5%
\<clipboard> \</clipboard> 87%

"Success rate" is measured by how often the AI correctly wraps information within the START_SEQ and END_SEQ tags when asked to copy to clipboard.

Results: [CLIPSTART] and [CLIPEND] is the most reliable style on Llama 3 8B. Consistently so over additional runs. The current defaults, with dashes (-), consistently perform the worst. I suspect dashes confuse the AI since they usually are used for bulleted lists and combined words.

ILikeAI commented 3 weeks ago

Dude this is big brain as hell!