puku

[puku @puku]

[2506.13284] AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy https://t.co/UlEbEMAmDC

[puku @puku]

roboflow/rf-detr: RF-DETR is a real-time object detection model architecture developed by Roboflow, SOTA on COCO and designed for fine-tuning. https://t.co/UQdCAqDXfc

[puku @puku]

Muon: An optimizer for hidden layers in neural networks | Keller Jordan blog https://t.co/eJavyZO2OD

[puku @puku]

[2410.21676] How Does Critical Batch Size Scale in Pre-training? https://t.co/jGQgAs8fRr

[puku @puku]

話題のGraphRAGとは - 内部構造の解析と実用性の考察 - アルファテックブログ https://t.co/TT2FolUfDA

[puku @puku]

[2408.07852] Training Language Models on the Knowledge Graph: Insights on Hallucinations and Their Detectability https://t.co/H4FV3gTikV

[puku @puku]

@rasbt In the llama technical report, it was mentioned that user prompts are not included in loss. Isn't it strange that prompts can be generated?🤔

[puku @puku]

@yonigottesman case-by-case https://t.co/aMAC593Ws4

[puku @puku]

Understanding in-context learning in transformers | ICLR Blogposts 2024 https://t.co/2pu8g052Tc

[puku @puku]

[2402.01035] 事前トレーニングとドメイン適応のためにトークナイザーを最大限に活用する https://t.co/mtm9kENaav

[puku @puku]

HQQクオンタイゼーション https://t.co/tLbSvqVh7t

[puku @puku]

Hybrid LLM: Cost-Efficient and Quality-Aware Query Routing | OpenReview https://t.co/oKhT0dV2Xa

[puku @puku]

[2402.01742] Towards Optimizing the Costs of LLM Usage https://t.co/CjJnuO5DcN

[puku @puku]

🤔🧐😲 https://t.co/6BIVZ81RWm

[Christian S. Perone @tarantulae]

Just sharing ~100 slides about PyTorch 2 internals focusing on recent innovations (Dynamo, Inductor, and ExecuTorch).
I had a lot of fun preparing this and hope you'll enjoy it. I'm planning to record it soon.

PDF: https://t.co/dlln8veNUA

Slideshare: https://t.co/c1OgpH0ZgC

[AK @_akhaliq]

Fast Llama 2 on CPUs With Sparse Fine-Tuning and DeepSparse by @neuralmagic with @Gradio demo

demo: https://t.co/YhFypHJSef

run with docker: https://t.co/zVXbiPcINf

duplicate space for private use: https://t.co/GHoGpBpJuh

blog: https://t.co/ZS0UAr6XlM

[Arena.ai @arena]

Introduce lookahead decoding:
- a parallel decoding algo to accelerate LLM inference
- w/o the need for a draft model or a data store
- linearly decreases # decoding steps relative to log(FLOPs) used per decoding step.

Blog: https://t.co/J4VvXgaBmm
Code: https://t.co/EgYNtFXkjH

[puku @puku]

https://t.co/TnX2IkDDg2

[Carlos E. Perez @IntuitMachine]

Introducing xVAL - the next evolution in numeric encoding for natural language processing!

Language models have achieved impressive results on textual data. But analyzing numbers - a fundamental building block of math, science, and life - remains a glaring blindspot. Existing encodings hamper models' ability to understand the relationships and patterns in numeric data. We wouldn't tokenize words letter-by-letter - why do we insist on discretizing numbers as strings of digits?

NVAL takes a radically different approach. It encodes the magnitude and meaning of numbers directly in the continuous embedding space. No more representing 1.2 and 1.3 as unrelated tokens! xVAL places nearby values nearby in the vector space, capturing the intrinsic mathematical meaning of numbers.

This simple but powerful shift in perspective unlocks game-changing advantages:

- Direct encoding of continuity and smoothness ➡️ Superior interpolation and generalization

- No more approximations from discreteness ➡️ Greater data efficiency

- Distance and proximity preserved ➡️ Numeric reasoning empowered

- Single token representation ➡️ Maximized computational performance

NVAL brings the benefits of continuity hard-coded into modern vision and audio transformers - now for language! It's the missing link for unlocking numeric reasoning in large models.

Applications are endless: scientific analysis, mathematical reasoning, financial forecasting, physics engines - anywhere numbers matter, xVAL delivers superior results with greater elegance and efficiency. The future is continuous!

[puku @puku]

JonathonLuiten/Dynamic3DGaussians https://t.co/AGlVevmSxo