Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
Starling-7B: Increasing LLM Helpfulness and Harmlessness with Rlaif (cs.berkeley.edu)
4 points by tosh on Nov 29, 2023 | hide | past | favorite | 2 comments


Starling-7B looks to be the new best of the tiny LLMs but it is a next-level hallucinator for things it doesn't know, by dint of being helpful to a fault I suspect.

>The phrase "to be railroaded" comes from the historical practice of "railroading" a train, which means to forcefully guide or manipulate the train's path. This practice was often used to get the train to stop at a specific station, even if the train's intended destination was elsewhere.

Another time it answered differently.

>The phrase is believed to have been derived from the practice of rushing a stolen train across state lines to avoid detection and prosecution.

That can't be right but sounds cool and plausible. Yi-34B gets closer to the current usage of the term given I don't think anyone knows where the exact real origin of "being railroaded" comes from anymore (land seizure, prisoner trains, working conditions, track metaphor).

>The phrase "to be railroaded" is an American English colloquialism that means to be treated unfairly or to be pushed into something against one's will, usually in a legal context. The origin of the phrase comes from the metaphor of being "run over" by a railroad, just as a train would run over anything in its path.

Apart from the leeway it grants itself to serve you I would say Starling-7B is very close to being competitive with Yi-34B because most answers are very similar.

Short live the new king in consumer LLM.


> Starling-7B-alpha scores 8.09 in MT Bench with GPT-4 as a judge, outperforming every model to date on MT-Bench except for OpenAI’s GPT-4 and GPT-4 Turbo.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: