I once had a pleasure of delving into the automotive mechanical engineering. Of course, most, if not all, materials ingested by OpenAI were obvious marketing straight from the brands website.
I started out the conversation multiple times anew, with explicit rules forbidding certain phrases. I couldn't make it stop throwing stuff like "best in class", "advanced", "sophisticated" no matter, what I did.
There will be demand for gpt's trained on an actual engineering material and it could actually be a huge gamechanger for that market.
That’s what I thought would happen. Actually, I thought it might be an easy way to get piles of text for LLM training. But we’d have to counter the bias or mostly use that one in highly-positive, enthusiastic applications. I did have a partial solution.
The author just deleted from the training data content with specific words likely to bias it. The test afterwards showed it worked. Reusing their concept, I think we could just remove or edit for honesty common words and phrases in marketing material. You’ve given some good examples.
We could also do that for “scientific” papers which oversell their results. Or anything else where what’s presented as certain is modified to say source(s) X claimed Y. Foundational materials, which trainers vet for quality, would get a lot more training runs before, during, and after riskier material.
I think there’s a lot of potential here by just trimming the fat out of otherwise useful documents. The LLM’s we build to support the work might also become great, lie detectors.
This seems to be some ChatGPT limitation. I also wanted it to omit certain phrases from the responses I tried to generate, but no amount of rules and explicit orders helped -- it would always include the same wording.
> There will be demand for gpt's trained on an actual engineering material and it could actually be a huge gamechanger for that market.
I imagine there will also be a lot of kinda-fraudulent supply from people who think: "I'll just take a cheap/commodity (badly) trained LLM, find just the right set of whack-a-mole prompts to make it appear to be making good output, and until customers catch-on the difference is pure profit."
Or perhaps they're open about it, and many customers just decide bad results cheap is better than premium data, which is... not a heartening thought.
eurekin|2 years ago
I started out the conversation multiple times anew, with explicit rules forbidding certain phrases. I couldn't make it stop throwing stuff like "best in class", "advanced", "sophisticated" no matter, what I did.
There will be demand for gpt's trained on an actual engineering material and it could actually be a huge gamechanger for that market.
nickpsecurity|2 years ago
Look at WizardLM Uncensored: https://www.reddit.com/r/LocalLLaMA/comments/1384u1g/wizardl...
The author just deleted from the training data content with specific words likely to bias it. The test afterwards showed it worked. Reusing their concept, I think we could just remove or edit for honesty common words and phrases in marketing material. You’ve given some good examples.
We could also do that for “scientific” papers which oversell their results. Or anything else where what’s presented as certain is modified to say source(s) X claimed Y. Foundational materials, which trainers vet for quality, would get a lot more training runs before, during, and after riskier material.
I think there’s a lot of potential here by just trimming the fat out of otherwise useful documents. The LLM’s we build to support the work might also become great, lie detectors.
distances|2 years ago
Terr_|2 years ago
I imagine there will also be a lot of kinda-fraudulent supply from people who think: "I'll just take a cheap/commodity (badly) trained LLM, find just the right set of whack-a-mole prompts to make it appear to be making good output, and until customers catch-on the difference is pure profit."
Or perhaps they're open about it, and many customers just decide bad results cheap is better than premium data, which is... not a heartening thought.