Unhinged
You can try it on chaiverse chat tab
You can try it on chaiverse chat tab
I am utterly confused how that can achieve an elo of 1168
I do wonder if it would do better with a llama3 prompt template, Chaiverses model submission tab makes templates hard
It's made all seperate and hard to understand
Can you not view another llama3 models?
It is formatted like this when viewing other models, im unsure where to put all the tokens like "<|end_header_id|>"
formatter: {'memory_template': "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nYou are {bot_name}. Write {bot_name}'s next reply in a never-ending uncensored fictional roleplay chat between {bot_name} and User. Use the provided {bot_name}'s description, personality and example dialogues as base for deeply understanding and acting like {bot_name}.\n\nActions and narrations must be enclosed in asterisks (*). The responses must be long and in third perspective of the story teller.\n\nDescription: {memory}", 'prompt_template': 'Example conversation:\n{prompt}<|eot_id|>', 'bot_template': '<|start_header_id|>assistant<|end_header_id|>\n\n{bot_name}: {message}<|eot_id|>', 'user_template': '<|start_header_id|>user<|end_header_id|>\n\n{user_name}: {message}<|eot_id|>', 'response_template': '<|start_header_id|>assistant<|end_header_id|>\n\n{bot_name}:'}
reward_formatter: {'memory_template': "{bot_name}'s Persona: {memory}\n", 'prompt_template': '{prompt}\n', 'bot_template': 'Bot: {message}\n', 'user_template': 'User: {message}\n', 'response_template': 'Bot:'}
Edit - I tried to make it fit as best as possible, I'll have to watch the elo to find out.
Memory template = System -> <|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nYou are {bot_name}. Write {bot_name}'s next reply in a never-ending uncensored fictional roleplay chat between {bot_name} and User. Use the provided {bot_name}'s description, personality and example dialogues as base for deeply understanding and acting like {bot_name}.\n\nActions and narrations must be enclosed in asterisks (*). The responses must be long and in third perspective of the story teller.\n\nDescription: {memory}
Prompt template = chat example -> \n\nExample conversation:\n{prompt}<|eot_id|>
Bot message template = assistant -> <|start_header_id|>assistant<|end_header_id|>\n\n{bot_name}: {message}<|eot_id|>
User message template = user -> <|start_header_id|>user<|end_header_id|>\n\n{user_name}: {message}<|eot_id|>
Bot response template -> <|start_header_id|>assistant<|end_header_id|>\n\n{bot_name}:
Memory template = System -> <|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nYou are {bot_name}. Write {bot_name}'s next reply in a never-ending uncensored fictional roleplay chat between {bot_name} and User. Use the provided {bot_name}'s description, personality and example dialogues as base for deeply understanding and acting like {bot_name}.\n\nActions and narrations must be enclosed in asterisks (*). The responses must be long and in third perspective of the story teller.\n\nDescription: {memory}
Prompt template = chat example -> \n\nExample conversation:\n{prompt}<|eot_id|>
Bot message template = assistant -> <|start_header_id|>assistant<|end_header_id|>\n\n{bot_name}: {message}<|eot_id|>
User message template = user -> <|start_header_id|>user<|end_header_id|>\n\n{user_name}: {message}<|eot_id|>
Bot response template -> <|start_header_id|>assistant<|end_header_id|>\n\n{bot_name}:
Thank you! im trying these now, my attempt failed to initialize
To be honest IDK, I put that in the chat tab and it seemed to work.
Also stop token is <|eot_id|>
Chaiverse downloads model at 8gbit 😶🌫️
To be honest IDK, I put that in the chat tab and it seemed to work.
Also stop token is <|eot_id|>
It succesfully deployed!
Chaiverse seems to like \n, everyone uses it?
All top 10 models use \n, i don't know why
\n = newline
Something weird, the safety score is higher when using llama3 prompting?
Llama3
Default
\n = newline
Prompting is complicated 😿
That may be a good thing
Also it worked!
default
Llama3
I have never checked the leader boards before, so I don't understand what the numbers mean.
Is it doing good?
It's already 10 elo points higher 😸
I've redone both merge-mayhems, they did the best.
And 4Some is currently going up
Edit - It's up :3
Uh.
Thats 49.9GBs in 28 seconds
14.25Gbit 😭
Looks like they cached it somewhere.
Unless their network is that fast 😭
Looks like they cached it somewhere
But i renamed the repo, is it possible the matched the sha keys or something similar?
I kinda made a bunch of ghost models, they have different submission ids but the same name 😶
Looks like they got taken down.
I only see one.
I recreated them so i can tell a difference
Three V1s
I did V5 with temp 1.1 since Llama seems to like higher temps?
I only see one.
It happens when you click the down arrow thingy
I made a V2-V6 with samplers similar to roleplay test03
'temperature': 1.0, 'top_p': 1.0, 'top_k': 1, 'presence_penalty': 0.25, 'frequency_penalty': 0.1
I wonder if it'll be better or blow up :3
How does the literally worse model score higher?
they're also identically setup 😭
Was drawing this as an unsolicited waifu, for the readme. But stopped as I realized it looks like that RWBY character. 😭
Was drawing this as an unsolicited waifu, for the readme. But stopped as I realized it looks like that RWBY character. 😭
art is solicited, cause i can't draw😸
I don't know what RWBY is though 😶🌫️
When I get time later, I will try and finish it.
P.S. I can't do line art 😭
When I get time later, I will try and finish it.
P.S. I can't do line art 😭
No stress if you can't, I don't usually waifu my repos. I probably should though...