Timeline for How do AI developers for LLM fix bugs when the LLM misbehaves, given that they cannot control what the LLM learns from the data?
Current License: CC BY-SA 4.0
23 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Dec 23, 2023 at 23:58 | vote | accept | curious | ||
| Dec 19, 2023 at 2:18 | comment | added | Ray | @JenserCube That answer says to tune the model. Tuning is just more training the model a bit more on specialized data, using a base model as a nice set of starting weights. If you don't like how the model works, then candied_orange's answer is the only correct one, because the model isn't actually buggy. A language model learns the distribution over text streams that is most likely to generate the corpus it was trained on. You don't want it to be stupid and racist? Don't train it on twitter. Anything you do after the fact is just a kludge that won't hold up to determined stress-testing. | |
| Dec 18, 2023 at 1:52 | comment | added | candied_orange | @Bergi correct. | |
| Dec 18, 2023 at 1:00 | comment | added | Bergi | @candied_orange The titular question is "How do AI developers for LLM fix bugs when the LLM misbehaves?", and your answer boils down to "You fix it with better data." | |
| Dec 17, 2023 at 23:34 | history | edited | candied_orange | CC BY-SA 4.0 |
added 178 characters in body
|
| Dec 17, 2023 at 22:29 | comment | added | candied_orange | “This states that changes cannot happen without changing the model” really? Where does it say that? | |
| Dec 17, 2023 at 22:15 | comment | added | JenserCube | This answer is misleading (at best). This states that changes cannot happen without changing the model, which we see on a daily basis when news articles mentions some problem. The right answer is softwareengineering.stackexchange.com/a/450153/431254. | |
| S Dec 17, 2023 at 18:56 | history | edited | candied_orange | CC BY-SA 4.0 |
English fixes
|
| S Dec 17, 2023 at 18:56 | history | suggested | tripleee | CC BY-SA 4.0 |
English fixes
|
| Dec 17, 2023 at 16:49 | review | Suggested edits | |||
| S Dec 17, 2023 at 18:56 | |||||
| Dec 17, 2023 at 16:41 | comment | added | Stef | Filtering the training data to remove racist training data is one way, but it's not the only way. During training, you could for instance add a racism penalty to the loss function, to encourage the network to give less racist output. | |
| Dec 17, 2023 at 15:01 | comment | added | candied_orange | @ChesterGillon better now? And yes. LLMs use random to choose which probabilistically weighted word comes next. Control the random etc. and it picks the same ones each time. | |
| Dec 17, 2023 at 14:57 | history | edited | candied_orange | CC BY-SA 4.0 |
added 5 characters in body
|
| Dec 17, 2023 at 12:51 | comment | added | Chester Gillon |
I'm not sure I understand the meaning of the and the random the output part of the answer. Did you mean and the random seeds, the output? I.e. does the LLM involve some sort of pseudo-random sequence which can cause the output to change?
|
|
| Dec 17, 2023 at 9:08 | comment | added | Kilian Foth | This is simply wrong. Every LLM that you experiment with has obvious limitations built in that you can circumvent by expressing your prompt in a way the publishers didn't think of. It is quite obvious that they can manipulate the permissible output rather than the input in a targeted way - just not perfectly. | |
| Dec 17, 2023 at 3:47 | history | edited | candied_orange | CC BY-SA 4.0 |
added 137 characters in body
|
| Dec 17, 2023 at 3:24 | comment | added | N. Virgo | yet in practise the companies that make these things are able to mitigate and control undesirable behaviour. Perhaps not perfectly but it works most of the time. There must be a huge box of tools they have available for this, including RLHF and probably all sorts of other techniques, none of which are mentioned in this answer. | |
| Dec 17, 2023 at 2:09 | comment | added | candied_orange | @whatsisname it’s called a frame challenge | |
| Dec 17, 2023 at 1:05 | comment | added | whatsisname | This reads like a non answer. You don't say how to control from surprising results that are non-obvious from the input. "Better data" is just "get gud" for AI | |
| Dec 16, 2023 at 20:56 | history | edited | candied_orange | CC BY-SA 4.0 |
deleted 36 characters in body
|
| Dec 16, 2023 at 20:30 | history | edited | candied_orange | CC BY-SA 4.0 |
edited body
|
| Dec 16, 2023 at 18:34 | comment | added | Greg Burghardt | They learn so quickly, too! The kids, I mean. The LLMs? Debatable. | |
| Dec 16, 2023 at 17:32 | history | answered | candied_orange | CC BY-SA 4.0 |