DeepSeekMath
PulseAugur coverage of DeepSeekMath — every cluster mentioning DeepSeekMath across labs, papers, and developer communities, ranked by signal.
2 day(s) with sentiment data
-
PPO vs DPO vs GRPO: Choosing the Right RL Method for Language Models
The choice between Proximal Policy Optimization (PPO), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO) for training language models depends on the available training signal. DPO is su…
-
GRPO: A memory-efficient RL method for LLM training
Group Relative Policy Optimization (GRPO) is a reinforcement learning technique that improves upon Proximal Policy Optimization (PPO) by eliminating the need for a separate critic model. Instead, GRPO compares multiple …
-
GRPO algorithm variant clarifies its role in RL fine-tuning
The GRPO algorithm, introduced by Shao and colleagues, is a variant of Proximal Policy Optimization (PPO) that modifies the training process by removing the critic network, which is computationally expensive. Instead of…
-
LLM Fine-Tuning Methods: SFT, LoRA, QLoRA, RFT, and Distillation Explained
The article outlines various methods for fine-tuning large language models, focusing on practical applications and tool choices. Supervised Fine-Tuning (SFT) is presented as a starting point, requiring labeled input-out…
-
Karpathy's nanochat uses simplified GRPO for RL loop
Andrej Karpathy's nanochat project includes a simplified reinforcement learning loop, labeled GRPO, that deviates from the standard GRPO algorithm. This loop uses a basic policy gradient method, essentially REINFORCE wi…
-
New RL policies boost high-frequency trading performance
Researchers have developed new reinforcement learning policies for high-frequency trading on limit order books. Their approach utilizes Order-Flow signals as a state representation and employs policy-gradient methods, s…
-
Open-source AntAngelMed model offers efficient medical AI with 103B parameters
Researchers have introduced AntAngelMed, a 103 billion parameter open-source medical language model. It utilizes a Mixture-of-Experts (MoE) architecture, activating only 6.1 billion parameters per query for enhanced eff…