Excellent results so far
Just wanted to say that by and large so far this model has performed better for me than any other to date.
Thank you so much!! :D
@Latent-Dreamscape What sorts of things are you using it for? What kind of hardware do you have it running on? You using oobabooger?
Great, glad it's working well for you. All credits to @timdettmers for the model's quality!
hi @TheBloke do you have a quick guide on using your 1-click docker? I tried pulling it down from the below link and it didn't work
@QuantumScribble running on a 4090 and using oobabooga for backend to sillytavern as well as running inference with oobabooga alone to generate stories and do some organizing tasks (like feeding it a list or document and asking it to organize some things, make lists, etc).
@TheBloke Indeed, @timdettmers is doing a wonderful job but I couldn't run them without you!
I think the praise is understated. This is the first local model that properly works for what I need it for. It doesn't know everything, but I can have long useful conversations with it, and it doesn't lose track of the topic. Just amazing. So far, I don't even see the need for a 65B, although it does seem to give terse replies at times.
I think the praise is understated. This is the first local model that properly works for what I need it for. It doesn't know everything, but I can have long useful conversations with it, and it doesn't lose track of the topic. Just amazing. So far, I don't even see the need for a 65B, although it does seem to give terse replies at times.
Can you share your workflow or maybe a character you like? I am still using GPT4x alpaca since last month as it's still been the best storytime/character based model so far outside of a jailbroken gpt4. This is good but for me it still comes off like an instructional model for any character stuff.
When I said "long conversations" I meant that I could talk to it in detail about project ideas and such. Haven't had a chance to use it for roleplaying, but it writes stories really well. Recently I used it to recommend a parts list to go with my rtx 3090. It's the only model that could pull it off without forgetting my requirements or getting stuck in some way. Most would refuse to update the parts list after a while when I requested changes.
yes this model seems gives (on subjective level) good responses compared to others. (In the models I have), after ChatGPT only this model answered correctly to "what is the second largest city to the capitol of England?". Also it seems I can make it to understand my context by conversation like a real person, i.e I started a question and finally, 'we' ended up in creating new question 'we' both agree :)
Great to hear people are getting good results.
hi @TheBloke do you have a quick guide on using your 1-click docker? I tried pulling it down from the below link and it didn't work
https://hub.docker.com/r/thebloke/runpod-pytorch-oneclick
@Co0ode Can you define "doesn't work?" I designed it for use on Runpod, but in theory it should work anywhere, assuming the Docker server is set up to pass through a GPU and network connectivity etc.
I have a problem, When I set the max_token to be relatively large (such as 512), the results of the model answer are not very accurate. For the same question, setting the max_token to a smaller value, such as 256, will have a much better effect. How to explain this?