Your models are too good.
I’ve been testing your model on Hugging Face’s free CPU, and I am genuinely impressed by its speed—it runs incredibly fast.
However, I’ve encountered a specific limitation regarding its reasoning. It seems the model is currently relying heavily on a predefined structure, almost like a hard-coded HTML response. For example, if we program it to respond to 'Hi' with a set phrase like 'How may I assist you?', it handles that perfectly. But if I ask anything outside of that specific context, the model struggles, gets stuck, or starts hallucinating.
Because I really value this model, I am currently fine-tuning this 50-million parameter version to improve its core reasoning and break it out of these rigid response patterns. I am putting a lot of effort into making it more flexible and capable of handling diverse queries.
I’d love to share the results of this fine-tuning with you once it’s ready. I hope you’ll like the improvements, and if you do, I’d be excited to continue working on this together. Looking forward to your thoughts!
Thats a good catch!
our first generation 50M model indeed does hallucinates! We recommend you fine-tune our 1.5 basemodel!
Okay 👍
@Hanuman2
Supra-1.5 went through reasoning tokens in PT. It will perform considerably better than Supra-1!
Sorry, i couldnt add that, I was in a hurry.
Also, you can try teachng it some light function calling. it went through OpenAi and anthropic tool calling tokens too. maybe small tasks like cheacking the wheather or sending amessage liek siri will be alright for this model!
Thanks for the update! It's great to hear that Supra-1.5 shows improvement with reasoning tokens. I'll definitely look into testing out the function calling capabilities for those smaller tasks you mentioned—that sounds like a perfect use case for it. Appreciate the tip!
I have done it Supra labs you can see your model here it is a demo space if you like
https://hanuman2-text.hf.space
It is for better than previous one and running smoothly on free space CPU
I have combined the qwen base to make it perfect and your model now it work fine and it knows how to call tools
Can you give me the demo code please?
FROM ubuntu:24.04
WORKDIR /app
RUN apt-get update && apt-get install -y curl libgomp1 && rm -rf /var/lib/apt/lists/*
RUN curl -L -o llama.tar.gz https://github.com/ggml-org/llama.cpp/releases/download/b8469/llama-b8469-bin-ubuntu-x64.tar.gz
&& tar -xzf llama.tar.gz
&& rm llama.tar.gz
&& find . -name ".so" -exec cp {} /app/ ;
&& find . -name "llama-server" -exec cp {} /app/ ;
&& chmod +x /app/llama-server
RUN curl -L -O https://huggingface.co/Hanuman2/Supralabs-Q4_K_M-GGUF/resolve/main/supralabs-q4_k_m.gguf
ENV LD_LIBRARY_PATH=/app
EXPOSE 7860
CMD ["./llama-server", "-m", "supralabs-q4_k_m.gguf", "--host", "0.0.0.0", "--port", "7860", "-c", "2048"]
the space you provided doesnt seem to be our model
Yes, you are correct that this is a fine-tuned version based on the Qwen architecture. I intentionally migrated to this base because the original Supra-50M model showed significant limitations in fundamental reasoning (e.g., failing at basic arithmetic like 2+2) and could not reliably execute function calling.
My objective was to ensure that Supra Labs' project transitions from a prototype to a production-ready system. By integrating your custom dataset with a more capable architecture, I have enabled reliable function calling and improved logical reasoning while maintaining your original data integrity.
I have attached screenshots of the baseline performance for comparison so you can see why this architectural shift was necessary for the project's success. This model is now fully optimized to handle diverse queries and tool-use tasks efficiently. I believe this provides a much higher value for Supra Labs, but I am happy to discuss the technical trade-offs further.
I am improving it
Okay I am fine tuning it more then I will show you result it's due to previous data
Alright bet!
also for more context, Supra models are sensitive to the hyperparameters. So please tune it as well!
You can now test it here please let me know if any discrepancy found
https://vedika-ai-vedika-ai2.hf.space
looks liek something a 350M model would say ngl
Did you like it
Yes! But this is not a SupraLabs model right? Because it has 0.5B params. Why did you call it SupraLabsthen? And why the logo?
i mean...ours is 50M, not 500M!
the space you provided doesnt seem to be our model
That's what I think too..
i got it. qwen finetune. sorry.
The reason the model parameters might appear different is due to the architectural optimization I performed. As I mentioned, the original 50M base was failing even on basic logic, so I migrated the fine-tuned Supra logic to a more capable Qwen base to ensure stability.
Regarding the naming and branding—my sole intention was to showcase the capabilities of the Supra Labs project to its full potential. I wanted to demonstrate how effectively the Supra dataset can perform when paired with a stable architecture. I apologize for any confusion caused by the naming; I will update the model card and repository description to explicitly state that this is a 'Community-optimized version' using the Supra Labs dataset.
I am fully committed to the Supra Labs project and my goal is simply to see it succeed. I hope this clarifies the intent, and I am happy to adjust the branding to align with your organization's guidelines.
It's okay I have read the message
Brothaaaaaa... that's basically Qwen's base model + finetuning.
I mean, by all means, good work, but it has no originality to it.
But either way, at SupraLabs, we are always positive about our community's work.
Keep it up!