Hacker News Reader

HNHacker News

Top New Best Ask Show Jobs

Comment by dartos | Hacker News Reader

Parent Full thread

dartos·

I wish there was a locked chatgpt version for researchers

View on HN

You could play with open models and different temperatures using the lmgen script from this repo we just released:

https://github.com/RichardKelley/hflm?tab=readme-ov-file#lmg...

there sort of is, if you install ollama (https://ollama.com) and then execute: ollama run llama2-uncensored it will install and run the local chat interface for llama2 in an uncensored version which gives a little bit better results with less guardrails. Same with wizardlm-uncensored and wizard-vicuna-uncensored. For reference, the SOTA version with guardrails you could run: ollama run llama3.1:405b

Some of the open-weights models come close to what you might want?

Sort of. They fill the “locked” requirement, but open models aren’t SOTA like gpt or Claud.

Llama3.1-405B benchmarks at a level that I would certainly consider comparable to other SOTA models, as long as multimodality isn’t needed.

Regardless, OpenAI provides access to quite a few of their older models through the API, since the API lets you pass in a specific model version. I’m sure the older models won’t be available forever, but that is a much more stable target for researchers than just opening the ChatGPT website and typing in things.

ChatGPT specifically is more than just a chat interface on top of got-4.

Their system prompt includes the current date and time among other information, making it very very hard to run reproducible experiments against it.

But it’s the tool most people are using.

LLAMA 3.1 405b-Instruct is state of the art at the time of writing. It averages slightly worse than ChatGPT 4o, but to a mostly negligible extent for most research.

Aside but -- it is also aligned to not generate profane text, right?

Abliterate it

Reply on news.ycombinator.com