I came across this new finetuned model based on Openchat 3.5 which is apparently trained used Reinforcement Learning from AI Feedback (RLAIF).
https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha
Check out this tweet: https://twitter.com/bindureddy/status/1729253715549602071
“Close to GPT4” is as true as “Me, Close to Usain bolt in the 100m dash” lol
Nope the research and proof is here not the parameters but the quality of data is the way my brotha
Woohoo
Was wondering how long this would take to show up.
That gap in coding is what makes me stay with GPT-4 until I don’t.
Have you tried DeepSeek, it’s pretty good at doing most things I’ve asked it to with Python.
It is pretty good, just not as good
“New RLAIF Finetuned 7b Model” Interesting. “beats Openchat 3.5” Nice! “and comes close to GPT-4” Bruh.
heheh i can’t read that any more… i really have become very prejudiced when comes to that… to be honest, when it comes to any comparison with GPT-4.
People have really to understand that even GPT-4 has been aligned, lobotomized and it has been massively downgraded in terms of its perfomance – due to security reasons (what is understandable for me), but anyway this thing still is an absolute beast. if we consider all the restrictions GPT-4 has to undergo, all the smartness at openAI, all the ressources at microsoft and so on, we have to realize that currently nothing is really comparable to GPT-4. Especially not 7B models.
I’ve seen the “… beats GPT-4” enough times that now whenever I see a title that suggests a tiny model can compete with GPT-4 I see it as a negative signal; that the authors are bullshitting through some benchmarks or some other shenanigans.
It’s annoying because the models might be legitimately good models for being open and within their weight class but now you’ve put my brain in BS detecting mode and I can’t trust you’ve done good faith measurement anymore.
Yeah I dont think authors are intentionally bullshitting or intentionally doing “benchmark cosmetics”, but maybe it’s more lack of knowledge on whats going on in terms of (most of) benchmarks and their the image that has become ruined in the meantime.
Sure, but name-dropping the biggest name in the game and comparing yourself favourably to it is a big swing. It’s either a naive at best marketing claim or it’s untrue.
There are SO many models “bullshitting through some benchmarks or some other shenanigans” that I’m cooking my own benchmark system LOL.
Yeah I just roll my eyes and continue onwards
Hard to believe but can’t wait to try.
Does somebody have a prompt template for this? Trying to run in ollama
Here’s what I’m using:
FROM starling-lm-7b-alpha.Q5_K_M.ggufPARAMETER stop <|end_of_turn|>PARAMETER stop <|im_sep|>TEMPLATE """GPT4 User: {{.Prompt}}<|end_of_turn|>GPT4 Assistant:"""how do you add your own gguf into ollama? it seems to be storing models as cryptic binary blobs in a folder.
generate the sha256 hash using sha256sum your_model.gguf
rename your_model.gguf to “sha256:_hash_” (replace _hash_ with the actual hash)
move it to /usr/share/ollama/.ollama/models/blobs folder
copy a manifest from a similar model in /usr/share/ollama/.ollama/models/
manifests/registry.ollama.ai/library and update the hash & filesize to match your model in the “image.model” entry.
repeat last step for the params entry
you can call the manifest folder/file whatever you like
Basically yes. https://github.com/jmorganca/ollama#import-from-gguf (download the gguf from huggingface eg https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha)
The first image posted; looks like it’s not even close to GPT-4?
Considering “close” as a relative word, it came closer than other open-source models. But you have a point too.
Do someone know why it writes the line feed code all the time in its answer ? <0 x 0 A>
Besides this, I find the model amazing.
Has been fixed in the unquantized model. They forgot to upload the tokenizer files https://twitter.com/banghuaz/status/1729375878612922724?s=12
https://huggingface.co/TheBloke/Starling-LM-7B-alpha-GGUF/discussions/1#65657dc79bf6665f10ebd941
Looks like TheBloke hasn’t picked it up. But then it has only been an hour 😂
It’s fixed now.
I was sceptical, but darn it’s good. Mistral is a fantastic base and with this technique these guys have pushed it another step closer. A lot of the answers I’m getting are on on par with old GPT-4 (pre-turbo, turbo in the API is a step up on old GPT-4 IMO).
the model can have a little of the test data as a treat
I can’t wait for the trustworthy closed sourced benchmarks. Can’t believe I’m saying that… but it’s honestly what we need.
Wonder if that’s a good startup idea? Something that can benchmark language models and charges a fee for doing so
Form huggingface model card,
Starling-RM-7B-alpha is a reward model trained from Llama2-7B-Chat.
From their webpage, https://starling.cs.berkeley.edu
Our reward model is fine-tuned from Llama2-7B-Chat
Yet, the model config.json
"max_position_embeddings": 8192, "model_type": "mistral", "num_attention_heads": 32, "num_hidden_layers": 32, "num_key_value_heads": 8, "rms_norm_eps": 1e-05, "rope_theta": 10000.0, "sliding_window": 4096,SO? Whoever is doing the PR has no f***ing idea what their student labors are actually doing.
What does it mean that an LLM is a reward model ? , I always thought of rewards only in the RL field . And how would the reward model be used during finetuning?
yeah I was put off by the lack of mention on the base model
If there is something somehow inherently superior about having a separate reward model, that should be teased out.
It would be nice to see stronger baselines / ablations for this reason. I realize it’s nigh impossible to keep up with the unrelenting pace of advances, so I don’t fault the authors here. That said, if there isn’t a compelling reason to keep the separate preference model, community people-hours will probably be best spent sticking with DPO/IPO to avoid the hyper-parameter tuning rabbit hole.
My guess: the way things are going, we’ll soon see a rough consensus emerge around a sane default DPO or Identity-PO recipe for fine-tunes (the same way we’ve seen gradual convergence around decoder-only transformer + rotational positional embeddings + group query attention + FlashAttention 2) to be applied absent a compelling reason to use a different reward signal.
No matter what, preference datasets like this are helpful. Pity about the license being claimed here, it’s hard to imagine it would hold up, but the specter is a bit of a hindrance.
rm is the reward model… not the same as the lm model. I tried the lm, wasn’t impressed. Gpt-3.5 did better for summarizing quotes. It was good, but I honestly think open hermes and or synthia 1.3b do better
How to earn VC money 101: “Beats GPT-4!”
And voila! you’re rich now.
And voila! You work for investors now.