Search Results for "strategy"

...It employs Multi-head Latent Attention (MLA) and the DeepSeekMoE architecture to enhance computational efficiency. The model introduces an auxiliary-loss-free load balancing strategy and a multi-token prediction training objective to boost performance. Trained on 14.8 trillion diverse, high-quality tokens, DeepSeek-V3 underwent supervised fine-tuning and reinforcement learning to fully realize its capabilities. Evaluations indicate that it outperforms other open-source models and rivals leading closed-source models, achieving this with a training duration of 55 days on 2,048 Nvidia H800 GPUs, costing approximately $5.58 million.

...Built on advanced deep learning architectures that combine Diffusion Transformer (DiT) and flow matching techniques, HY-Motion scales these approaches to the billion-parameter level, resulting in strong instruction-following capabilities and richer motion outputs compared to existing open-source models. The training strategy for the HY-Motion series includes extensive pre-training on thousands of hours of varied motion data, fine-tuning on curated high-quality datasets, and reinforcement learning with human feedback, which improves both the plausibility and adaptability of generated motion sequences.

MedicalGPT training medical GPT model with ChatGPT training pipeline, implementation of Pretraining, Supervised Finetuning, Reward Modeling and Reinforcement Learning. MedicalGPT trains large medical models, including secondary pre-training, supervised fine-tuning, reward modeling, and reinforcement learning training.

...The project demonstrates a system that orchestrates large language models (LLMs) — like those from major providers — with carefully engineered prompting, reasoning workflows, and dynamic strategies, to tackle the abstract, logic-heavy problems in ARC-AGI. Instead of relying on a single prompt or fixed strategy, their solver dynamically adapts the reasoning path, selecting what to ask or analyze next depending on intermediate results — effectively compositing reasoning, perception, and program synthesis (or symbolic manipulation) in a loop. The repository allows others to reproduce their results, experiment with different LLM backends (e.g. the user may supply keys for supported models), and observe how their adaptive meta-system handles the logic and abstraction challenges.

...Instead of relying only on static training data, the agent performs searches, visits pages, and extracts evidence before forming answers to complex queries. It is built to operate end-to-end: planning a research strategy, gathering sources, reasoning over conflicting claims, and writing a grounded response. The repository includes evaluation results on multi-step QA and research benchmarks, illustrating how web-time context boosts accuracy. Because the system is modular, you can swap the search component, reader, or policy to fit private deployments or different data domains. ...

VibeThinker is a compact but high-capability open-source language model released by WeiboAI (Sina AI Lab). It contains about 1.5 billion parameters, far smaller than many “frontier” models, yet it is explicitly optimized for reasoning, mathematics, and code generation tasks rather than general open-domain chat. The innovation lies in its training methodology: the team uses what they call the Spectrum-to-Signal Principle (SSP), where a first stage emphasizes diversity of reasoning paths (the...

DeiT (Data-efficient Image Transformers) shows that Vision Transformers can be trained competitively on ImageNet-1k without external data by using strong training recipes and knowledge distillation. Its key idea is a specialized distillation strategy—including a learnable “distillation token”—that lets a transformer learn effectively from a CNN or transformer teacher on modest-scale datasets. The project provides compact ViT variants (Tiny/Small/Base) that achieve excellent accuracy–throughput trade-offs, making transformers practical beyond massive pretraining regimes. Training involves carefully tuned augmentations, regularization, and optimization schedules to stabilize learning and improve sample efficiency. ...