Qwen3-4B-SafeRL is a safety-aligned version of the Qwen3-4B model. It enhances the model's robustness against harmful or adversarial prompts through reinforcement learning training. This version uses a hybrid reward function for optimization, balancing three objectives: safety, usefulness, and minimizing unnecessary rejections.

Natural Language ProcessingTransformersEnglish

G²RPO is a novel reinforcement learning framework specifically designed for preference alignment in flow models, significantly improving the generation quality through a granular reward evaluation mechanism.

MultimodalDiffusers

Qwen3-4B-SafeRL is a safety-aligned version based on the Qwen3-4B model. It is trained through reinforcement learning and combined with the reward signals of Qwen3Guard-Gen, enhancing the model's robustness against harmful or adversarial prompts. While ensuring safety, it avoids overly simple or evasive rejection behaviors.

Natural Language ProcessingTransformers

Qwen3-4B-I-1509 is a dedicated model fine-tuned based on the Qwen3-4B-Instruct-2507 base model, focusing on tool usage and function call generation tasks. This model is trained using a reinforcement learning method with multi-signal rewards and demonstrates good performance in multiple evaluation benchmarks.

Natural Language ProcessingTransformersEnglish

SRPO is a human preference alignment method for diffusion models. Through Direct-Align technology and semantic relative preference optimization, it significantly improves the realism and aesthetic quality of the FLUX.1-dev model, and solves the problems of high computational cost in multi-step denoising and dependence on offline reward fine-tuning.

Computer VisionDiffusers

DeepSeek-GRM-16B is a generative reward model based on Self-Principled Critique Tuning (SPCT), which can generate a transparent 'Principle → Critique → Score' evaluation process for query-responses and can be used in tasks such as reinforcement learning, evaluation, and data collection of large language models.

Natural Language ProcessingTransformersMultiple Languages

Seed-X-RM-7B is a reward model in the Seed-X series, specifically designed to evaluate translation quality. Based on the 7-billion-parameter Mistral architecture, this model can assign reward scores to multilingual translations and supports the evaluation of translation quality among 25 languages.

Natural Language ProcessingSafetensors

This is a quantized version of the judgment model with 32B parameters, suitable for tasks such as model evaluation, reward modeling, and text sorting.

Natural Language ProcessingTransformersEnglish

The static quantization version of CompassJudger-2-32B-Instruct, suitable for tasks such as model judgment, evaluation, reward modeling, and text sorting.

Natural Language ProcessingTransformersEnglish

POLAR-7B is a scalar reward model based on large-scale pretraining. It adopts an innovative policy discriminative learning paradigm and can effectively distinguish policies and align with human preferences.

Natural Language ProcessingTransformersMultiple Languages

GTA1 is a GUI positioning model based on Reinforcement Learning (GRPO). It achieves precise positioning by directly rewarding successful clicks, avoiding lengthy thought-chain reasoning.

MultimodalTransformers

Llama-3.1-8B-Instruct-RM-RB2 is one of the reward models released by the Allen Institute for Artificial Intelligence. It is fine-tuned based on Llama-3.1-8B-Instruct and is used to evaluate and optimize the preference learning of generative models.

Natural Language ProcessingTransformersEnglish

Llama-3.1-70B-Instruct-RM-RB2 is one of the 7 reward models released by the Allen Institute for Artificial Intelligence, used to develop benchmarks and conduct correlation analysis with downstream PPO/Best-of-N performance.

Natural Language ProcessingTransformersEnglish

ThinkPRM-7B is a generative process reward model based on the R1-Distill-Qwen-7B architecture, specifically designed for step-by-step verification of inference processes. It evaluates the correctness of each step in reasoning tasks such as mathematical problem-solving and code generation by generating explicit verification thought chains, featuring high data efficiency and strong interpretability.

Natural Language ProcessingTransformersOther

RM-R1 is a training framework for reasoning reward models (ReasRM), which evaluates candidate answers by generating scoring criteria or reasoning trajectories, providing interpretable evaluations.

Natural Language ProcessingTransformersEnglish

RM-R1 is a training framework for reasoning reward models (ReasRM), which evaluates candidate answers by generating scoring criteria or reasoning traces, providing explainable judgments.

Natural Language ProcessingTransformersEnglish

RM-R1 is a training framework for inference reward models. It evaluates two candidate answers by generating structured scoring criteria or inference trajectories and provides interpretable reasons.

Natural Language ProcessingTransformersEnglish

RM-R1 is a training framework for reasoning reward models (ReasRM), which evaluates candidate answers by generating scoring criteria or reasoning traces, significantly improving accuracy and interpretability compared to traditional reward models.

Natural Language ProcessingTransformersEnglish

RM-R1 is a training framework for reasoning reward models (ReasRM), which evaluates candidate answers by generating scoring criteria or reasoning traces, providing explainable assessments.

Natural Language ProcessingTransformersEnglish

RM-R1 is a framework for reward modeling through reasoning trajectory generation, offering significant improvements in accuracy and interpretability compared to traditional methods

Natural Language ProcessingTransformersEnglish