Spaces:
Running
Lousy quality of the trained models?
Guys, why is the output of these models so bad? The claim was GPT2 like performance at least right? Below is the output of the trained models uploaded here (in the demo space too) and similar reports on X.
Base-20:
Question: What is the capital of Belgium?
<|bos|>The capital of Belgium is 11,000 square kilometers and is located in the north of the country. The capital of Belgium is Brussels. The capital of Belgium is located in the north of the country. The capital of Belgium is located in the north of the country. The
Chat-20:
Always answers this: The first step is to identify the key elements of your story. This involves understanding the core elements that will drive your narrative forward, such as characters, plot twists, and themes. Think about the kind of story you want to tell and what kind of themes you want to explore.
Once you have a sense of the overall narrative arc, you can start to develop your characters. What are their motivations, desires, and fears? What are their strengths and weaknesses? How do they change and grow throughout the story? Consider their backstories, relationships, and conflicts to add depth and nuance to your characters.
The plot is the backbone of your story, and it's essential to create a clear structure that flows logically. Think about the key events that will drive the narrative forward, and how they relate to one another. Consider the pacing, tension, and conflict resolution to keep your readers engaged.
As you develop your characters and plot, you can start to think about the themes you want to explore. What messages or ideas do you want to convey through your story? What commentary do you want to leave on your readers? By answering these questions, you'll be able to create a cohesive and compelling narrative that resonates with your audience.
Finally, consider the tone and atmosphere you
Something is wrong with either the inference code or the model. I'd imagine they are working on it. To get an idea of how it works in the meantime, feel free to try https://huggingface.co/spaces/sdobson/nanochat
Yours is much better! Did you do some changes?
I took a different approach for inference. My space uses pytorch directly and runs on CPU.
This one uses the huggingface library and runs on GPU. Very cool, but I think not quite there yet.
thanks for sharing @sdobson .Yeah I'm still wrestling with the inference implementation in transformers to get the most out of the model. Using the original implementation, performance is as expected.
also. @sdobson feel free to join the org and duplicate your space into this org if you want to share it.
Sure, happy to share. Here's the space: https://huggingface.co/spaces/nanochat-students/nanochat-100dollar
This is a great initiative, by the way - awesome job!
@sdobson Nice!
I also finished the port to transformers for this space: https://huggingface.co/spaces/nanochat-students/chat-d20-demo . So happy to hear your feedback on the model performance.
now it's in transformers, it should unlock loads of downstream tasks
@burtenshaw Wow - loving the speed on GPU! Had a quick play and responses look good. Would be cool if it could stream responses token-by-token as they're generated. Also, some sampling would be nice to get some variety in outputs.