Small models don't "know" as much so they hallucinate more. They are better suited for generations that are based in a ground truth, like in a RAG setup.
A better comparison might be Flash 2.0 vs 4o-mini. Even then, the models aren't meant to have vast world knowledge, so benchmarking them on that isn't a great indicator of how they would be used in real-world cases.
It's for visualizing datasets where fine-grained cluster details and broader relationships matter. There are example renderings in the paper.
From the README:
> PaCMAP (Pairwise Controlled Manifold Approximation) is a dimensionality reduction method that can be used for visualization, preserving both local and global structure of the data in original space. PaCMAP optimizes the low dimensional embedding using three kinds of pairs of points: neighbor pairs (pair_neighbors), mid-near pair (pair_MN), and further pairs (pair_FP).
> Previous dimensionality reduction techniques focus on either local structure (e.g. t-SNE, LargeVis and UMAP) or global structure (e.g. TriMAP), but not both, although with carefully tuning the parameter in their algorithms that controls the balance between global and local structure, which mainly adjusts the number of considered neighbors. Instead of considering more neighbors to attract for preserving glocal structure, PaCMAP dynamically uses a special group of pairs -- mid-near pairs, to first capture global structure and then refine local structure, which both preserve global and local structure. For a thorough background and discussion on this work, please read our paper.
If you look in the `config.json`[1] it shows `Zamba2ForCausalLM`. You can use a version of the transformers library to do inference that supports that.
The model card states that you have to use their fork of transformers.[2]
This is true. Devs are looking for frameworks. See CrewAI who refuses to allow users to disable some pretty aggressive telemetry, yet they have a huge number of GH stars.
The abstractions are handy if you have no idea what you are doing but it's not groundbreaking tech.
> CoreNet evolved from CVNets, to encompass a broader range of applications beyond computer vision. Its expansion facilitated the training of foundational models, including LLMs.
It looks like a mid-level implementations of training and inference. You can see in their "default_trainer.py"[1] that the engine uses Tensors from torch but implements its own training method. They implement their own LR scheduler and optimizer; the caller can optionally use Adam from torch.
It's an interesting (maybe very Apple) choice to build from the ground up instead of partnering with existing frameworks to provide first class support in them.
There's gotta be somewhere in the middle. Vercel's movements feel a lot like the "Embrace, extend, and extinguish" playbook.
Maybe there is a class of developer out there that doesn't get spooked by that but it definitely has created an adversarial place for Vercel in my mind. I feel like I need to be careful when touching anything Vercel have touched so that I don't fall into a trap.
Llama.cpp is an inference engine. The author of llama.cpp designed gguf. Funcionary is a model that does function calling. You can download functionary weights in the gguf format and then run it using llama.cpp on low-end machines using CPU or GPU or a mix of both.
This paper is well written. The results are pretty wild. They observed some amazing reduction in training resources required to achieve similar benchmarks to models trained on conventional data:
> We observe that even at the first checkpoint (10B tokens) of WRAP training, the average perplexity of the LLM on the Pile is lower than that achieved by pre-training on C4 for 15 checkpoints. This suggests a 15x pre-training speed-up.
A better comparison might be Flash 2.0 vs 4o-mini. Even then, the models aren't meant to have vast world knowledge, so benchmarking them on that isn't a great indicator of how they would be used in real-world cases.