Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
0
7.93k
BLIP3-o: A Family of Fully Open Unified Multimodal Models-Architecture, Training and Dataset Jiuhai Chen1,2* Zhiyang Xu3* Xichen Pan4* Yushi Hu5* Can Qin1, Tom Goldstein2, Lifu Huang6, Tianyi Zhou2, Saining Xie4, Silvio Savarese1 Le Xue1†, Caiming Xiong1‡, Ran Xu1‡ 1Salesforce Research 2University of Maryland, 3Virgini...
Contents Introduction Unified Multimodal for Image Generation and Understanding 2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Combining Autoregressive and Diffusion Models . . . . . . . . . . . . . . . . . . Image Generation in Unified Multimodal 3.1 Image Encoding and...
A young woman with freckles wearing a straw hat. [IMG] Autoregressive Model Diffusion Transformer Flow Matching Understanding Generation Text Embedding Visual Feature Gaussian Noise CLIP CLIP Text Token Cross Entropy A cheerful Corgi standing behind a camera ... Figure 1: The architecture of BLIP3-o. For image understa...
Figure 2: Visualization results of BLIP3-o 8B at 1024×1024 resolution.
foundation model research like BLIP-3 , we fully open-source our models including model weights, code, pretraining and instruction-tuning datasets, and evaluation pipelines. We hope that our work will support the research community and drive continued progress in the unified multimodal domain. Unified Multimodal for Im...
autoregressive model generates visual features, how do we optimally align them with the ground-truth image features, or more generally, how should we model the distribution of these continuous visual features: via a simple MSE loss, or by employing a diffusion-based approach? Thus, we conduct a comprehensive exploratio...
3.2 Modeling Latent Image Representation After obtaining continuous image embeddings, we proceed to model them using autoregressive architectures. Given a user prompt (e.g., "A young woman with freckles wearing a straw hat."), we first encode the prompt into a sequence of embedding vectors C using the autoregressive mo...
Prompt Visual Feature Autoregressive Model (a) CLIP + MSE MSE CLIP Autoregressive Model Flow Matching (b) CLIP + Flow Matching Diffusion Transformer CLIP (c) VAE + Flow Matching VAE Autoregressive Model Flow Matching Diffusion Transformer Figure 3: Three design choices for image generation in unified multimodal model. ...
Results We report the FID score on MJHQ-30k for visual aesthetic quality, along with GenEval and DPG-Bench metrics for evaluating prompt alignment. We plot the results for each design choice at approximately every 3,200 training steps. Figure 4 shows that CLIP + Flow Matching achieves the best prompt alignment scores o...
1st Stage: Image Understanding Text 2nd Stage: Image Generation (b) Sequential Training LLM CLIP Image Text MLLM Visual Feature Text Embedding Image CLIP Diffusion Transformer Flow Matching LLM Visual Feature Text CLIP Image Text (a) Joint Training Image Diffusion Transformer Flow Matching CLIP Text Embedding Figure 5:...
normalization (RMSNorm before and after attention/MLP) and Grouped-Query Attention to enhance stability and reduce computation. Based on empirical results, this architecture achieves fast, highquality generation. 5.2 Training Recipe Stage 1: Pretraining for Image Generation For 8B model, we combine around 25 million op...
Model VQAv2 MMBench SEED MM-Vet MME-P MME-C MMMU RWQA TEXTVQA EMU2 Chat 34B 62.8 48.5 34.1 66.6 Chameleon 7B 19.8 27.2 8.3 202.7 22.4 39.0 0.0 Chameleon 34B 32.7 9.7 604.5 38.8 39.2 0.0 Seed-X 17B 63.4 70.1 66.5 43.0 1457.0 35.6 VILA-U 7B 79.4 66.6 57.1 33.5 1401.8 32.2 46.6 48.3 LMFusion 16B 72.1 1603.7 367.8 41.7 60....
feeding images into the image understanding vision encoder and then reconstructing them via the image generation model, to seamlessly bridge image understanding and generation. Building on this capability, we will collect instruction tuning datasets to adapt the model to various downstream applications. Related Work Re...
• A giant glass bottle filled with a miniature summer forest inside. • Walk through of frozen streets of Manhattan, New York City-frozen trees and a frozen Empire State Building. • A lighthouse standing alone in a stormy sea • A lone wolf beneath shimmering northern lights. • A glowing deer walking through a neon-lit f...
Flow-GRPO: Training Flow Matching Models via Online RL Jie Liu1,3,5∗Gongye Liu2,3* Jiajun Liang3 Yangguang Li1 Jiaheng Liu4 Xintao Wang3 Pengfei Wan3 Di Zhang3 Wanli Ouyang1,5† 1MMLab, CUHK 2Tsinghua University 3Kling Team, Kuaishou Technology 4Nanjing University 5Shanghai AI Laboratory jieliu@link.cuhk.edu.hk wlouyang...
(b) Image Quality (c) Preference Score (a) GenEval Performance Figure 1: (a) GenEval performance rises steadily throughout Flow-GRPO's training and outperforms GPT-4o. (b) Image quality metrics on DrawBench remain essentially unchanged. (c) Human Preference Scores on DrawBench improves after training. Results show that...
Related Work RL for LLM. Online RL has effectively improved the reasoning abilities of LLMs, such as DeepSeek-R1 and OpenAI-o1 , using policy gradient methods like PPO or value-free GRPO . GRPO is more memory efficient by removing the need for a value network, so we adopt it in this work. PPO can also be applied to flo...
... ... ... ... ... Flow Matching T2I Model SDE Sampling Prompt: A photo of four cups. Group Relative Policy Optimization Reward Function JFlow−GRPO= Jclip−βDKLπθ | πθold πθatst s0 ∼N(0, I) sT Policy Optimization R1 R2 R3 RG R4 Group Computation መA1 መA2 መA3 መAG መA4 Rewards Advantages ... ... Denoising Reduction (T = 10...
forward process adds Gaussian noise step by step, and the reverse process approximates a score-based SDE solver via a Markov chain with decreasing variance. In contrast, flow matching models use a deterministic ODE for the forward process: dxt = vtdt, (6) where vt is learned via the flow matching objective in Eq. 2. A ...
Compositional Image Generation. GenEval assesses T2I models on complex compositional prompts-like object counting, spatial relations, and attribute binding-across six difficult compositional image generation tasks. We use its official evaluation pipeline, which detects object bounding boxes and colors, then infers thei...
Table 1: GenEval Result. Best scores are in blue , second-best in green . Results for models other than SD3.5-M are from or their original papers. Obj.: Object; Attr.: Attribution. Model Overall Single Obj. Two Obj. Counting Colors Position Attr. Binding Diffusion Models LDM 0.37 0.92 0.29 0.23 0.70 0.02 0.05 SD1.5 0.4...
Table 2: Performance on Compositional Image Generation, Visual Text Rendering, and Human Preference benchmarks, evaluated by task performance on test prompts, and by image quality and preference scores on DrawBench prompts. ImgRwd: ImageReward; UniRwd: UnifiedReward. Model Task Metric Image Quality Preference Score Gen...
Ours (W/o KL) SD-3.5-M A photo of four apples a futuristic spaceship cockpit with a digital screen displaying " fuel low " warning, set against the backdrop of a star - studded galaxy, with control panels and futuristic interfaces around. Abraham Lincoln gives an outdoor speech to large crowd Ours (W/o KL) Quality Degr...
significant gains on T2I-CompBench++ . This comprehensive benchmark for open-world compositional T2I generation features object classes and relationships substantially different from our model's GenEval-style training data. Table 3: T2I-CompBench++ Result. This evaluation uses the same model presented in Table 1, which...
MUON IS SCALABLE FOR LLM TRAINING TECHNICAL REPORT Jingyuan Liu1 Jianlin Su1 Xingcheng Yao2 Zhejun Jiang1 Guokun Lai1 Yulun Du1 Yidao Qin1 Weixin Xu1 Enzhe Lu1 Junjie Yan1 Yanru Chen1 Huabin Zheng1 Yibo Liu1 Shaowei Liu1 Bohong Yin1 Weiran He1 Han Zhu1 Yuzhi Wang1 Jianzhou Wang1 Mengnan Dong1 Zheng Zhang1 Yongsheng Kan...
Muon is Scalable for LLM Training TECHNICAL REPORT Introduction The rapid advancement of large language models (LLMs) (OpenAI et al. 2024; DeepSeek-AI et al. 2024; Grattafiori et al. 2024; Gemini Team et al. 2024) has significantly pushed forward the progress in artificial general intelligence. However, training capabl...
Muon is Scalable for LLM Training TECHNICAL REPORT Newton-Schulz Iterations for Matrix Orthogonalization Equation 1 is calculated in an iterative process. At the beginning, we set X0 = Mt/∥Mt∥F. Then, at each iteration k, we update Xk from Xk−1 as follows: Xk = aXk−1 + b(Xk−1XT k−1)Xk−1 + c(Xk−1XT k−1)2Xk−1 (2) where X...
Muon is Scalable for LLM Training TECHNICAL REPORT 10000 20000 30000 40000 50000 60000 Training Iterations 2.25 2.30 2.35 2.40 2.45 2.50 2.55 Validation Loss 10000 20000 30000 40000 50000 60000 Training Iterations 0.020 0.015 0.010 0.005 0.000 0.005 0.010 0.015 Difference in Val Loss (Vanilla Muon - Muon w/ weight deca...
Muon is Scalable for LLM Training TECHNICAL REPORT Algorithm 1 Distributed Muon Require: Full Gradients G, DP partitioned Momentum m, DP partitioned parameters p, momentum μ. 1: // Reduce-scatter G on DP for correct gradients 2: g = reduce_scatter(G, dp_group) 3: // Apply momentum to g using local partitioned momentum ...
Muon is Scalable for LLM Training TECHNICAL REPORT Table 1: Controlling Muon's Update RMS Across Different Model Params Methods Training loss Validation loss query weight RMS MLP weight RMS Baseline 2.734 2.812 3.586e-2 2.52e-2 Update Norm 2.72 2.789 4.918e-2 5.01e-2 Adjusted LR 2.721 2.789 3.496e-2 4.89e-2 Experiments...
Muon is Scalable for LLM Training TECHNICAL REPORT Table 2: Scaling Law Models and Hyper-Parameters # Params. w/o Embedding Head Layer Hidden Tokens LR Batch Size* 399M 8.92B 9.503e-4 545M 14.04B 9.143e-4 822M 20.76B 8.825e-4 1.1B 28.54B 8.561e-4 1.5B 38.91B 8.305e-4 *In terms of number of examples in 8K context length...
Muon is Scalable for LLM Training TECHNICAL REPORT 2. 33B to 5.2T tokens: In this stage, the learning rate decays from 4.2e-4 to 4.2e-5 in a cosine style. We keep the batch size at 2048 until 200B tokens, and then doubled to 4096 for the remaining; 3. 5.2T to 5.7T tokens: In this stage (also referred as the cooldown st...
Muon is Scalable for LLM Training TECHNICAL REPORT Table 5: Comparison of different models on various benchmarks. Benchmark (Metric) Llama3.2-3B Qwen2.5-3B DSV2-Lite Moonlight Activated Param† 2.81B 2.77B 2.24B 2.24B Total Params† 2.81B 2.77B 15.29B 15.29B Training Tokens 9T 18T 5.7T 5.7T Optimizer AdamW Unknown AdamW ...
Muon is Scalable for LLM Training TECHNICAL REPORT 0.80 0.85 0.90 AttnQO 0.85 0.90 0.95 AttnKV 0.93 0.94 0.95 Experts 0.90 0.92 0.94 SharedExperts 0.7 0.8 0.9 Router 0.95 0.96 0.97 0.98 Dense Training Iterations (K) SVD Entropy AdamW Muon Figure 4: SVD entropy of weight matrices across different training iterations. We...
Muon is Scalable for LLM Training TECHNICAL REPORT Discussions There are several possible directions for future research that could further explore and expand upon the current findings. Incorporating All Parameters into the Muon Framework Currently, the Muon optimizer is utilized in conjunction with the Adam optimizer,...
Muon is Scalable for LLM Training TECHNICAL REPORT
Reinforcement Learning for Reasoning in Large Language Models with One Training Example Yiping Wang1 † ∗ Qing Yang2 Zhiyuan Zeng1 Liliang Ren3 Liyuan Liu3 Baolin Peng3 Hao Cheng3 Xuehai He4 Kuan Wang5 Jianfeng Gao3 Weizhu Chen3 Shuohang Wang3 † Simon Shaolei Du1 † Yelong Shen3 † 1University of Washington 2University of...
Steps Accuracy (%) MATH 500 (avg@1) 1.2k DSR-sub 7.5k MATH train set 2 shot { 13} 1 shot { 13} 1 shot { format reward Steps 17.5 20.0 22.5 25.0 27.5 30.0 32.5 35.0 37.5 Accuracy (%) Average on 6 benchmarks 1.2k DSR-sub 7.5k MATH train set 2 shot { 13} 1 shot { 13} 1 shot { format reward Figure 1: RLVR with 1 example (g...
example, overfitting does not occur until after approximately 1.4k training steps. Even post-overfitting, while the model's reasoning outputs for the training example become incomprehensible multilingual gibberish mixed with correct solutions, its test performance remains strong, and the reasoning outputs for the test ...
We then select examples according to this straightforward ranking criterion. For instance, π1, identified by the historical variance score on Qwen2.5-Math-1.5B, performs well in 1-shot RLVR (Sec. 3.2.3, 3.3). We also choose additional examples from diverse categories among {π1, . . . , π17} and evaluate them under 1-sh...
Table 2: Example π1. It is selected from DSR-sub (Sec. 3.1). Prompt of example π1: The pressure \\( P \\) exerted by wind on a sail varies jointly as the area \\( A \\) of the sail and the cube of the wind's velocity \\( V \\). When the velocity is \\( 8 \\) miles per hour, the pressure on a sail of \\( 2 \\) square fe...
Problem: As shown in Table 2 Answer: 12.8 Problem: Find the roots of $(x - 3)^3 + (x -7)^3 = (2x - 10)^3.$ Answer: 3, 5, 7 Training Response Evaluation Response (MATH500) Step 0: "Step 1:...\\[ P = kAV^3\\] ...Step 2: ...\\[k = \\frac{4}{1024} = \\frac{1}{256} \\]... Step 3: ..\\[ V = \\sqrt {2048} \\]\n\n\\[ V = 12.88...
Table 3: 1(Few)-Shot RLVR performance (%) for different categories in MATH500. Here for MATH500, we consider Algebra (Alg.), Count & Probability (C.P.), Geometry (Geo.), Intermediate Algebra (I. Alg.), Number Theory (N. T.), Prealgebra (Prealg.), Precalculus (Precal.), and MATH500 Average (Avg.). We report the best mod...
Table 4: 1(few)-shot RLVR is viable for different models and RL algorithm. "Random" denotes the 16 examples randomly sampled from 1.2k DSR-sub. Format reward (Appendix C.2.3) serves as a baseline for format correction. More details are in Appendix C.1, and we also include the results of Qwen2.5-Math-1.5B-Instruct and Q...
Table 5: Ablation study of loss function and label correctness. Here we use Qwen2.5-Math-1.5B and example π1. "+" means the component is added. "Convergence" denotes if the training accuracy saturates (e.g. Fig. 2). "-0.003" is the coefficient of entropy loss (default -0.001). We report the best model performance on ea...
the temperature to t = 1.0 yields an additional 0.8% gain. More discussions about entropy loss and post-saturation generalization are in Appendix C.2.2. 4.2 Entropy-Loss-Only Training & Label Correctness Table 6: Training with only entropy loss using π1 can partially improve base model performance, but still perform wo...
Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning Bowen Jin1, Hansi Zeng2, Zhenrui Yue1, Jinsung Yoon3, Sercan ̈O. Arık3, Dong Wang1, Hamed Zamani2, Jiawei Han1 1 Department of Computer Science, University of Illinois at Urbana-Champaign 2 Center for Intelligent Information Retr...
approaches offer greater adaptability but are difficult to scale effectively due to their reliance on large-scale, high-quality annotated trajectories and the inherent non-differentiability of the search operation, which renders end-to-end gradient descent-based optimization inapplicable (Schick et al., 2023; Asai et a...
Related Works 2.1 Large Language Models and Retrieval Despite demonstrating remarkable reasoning (Guo et al., 2025) and coding (Guo et al., 2024) capabilities, LLMs (Zhao et al., 2023; Team, 2024; Achiam et al., 2023) often lack domainspecific knowledge (Peng et al., 2023; Li et al., 2023) and are prone to hallucinatio...
q o Reward Model Value LLM Reference LLM r v A GAE q Reward Model Reference LLM Group Computation o1 o2 oG ... r1 r2 rG ... A1 A2 AG ... Policy LLM Search Engine Rollout Module Policy LLM Search Engine Rollout Module KL Trained Models Frozen Model Search Engine PPO w. Search Engine GRPO w. Search Engine Figure 1: Demon...
GRPO with Search Engine. To improve policy optimization stability and avoid the need for an additional value function approximation, Group Relative Policy Optimization (GRPO) is introduced in Shao et al. (2024). GRPO differs from PPO by leveraging the average reward of multiple sampled outputs as a baseline rather than...
Algorithm 1 LLM Response Rollout with Multi-Turn Search Engine Calls Require: Input query x, policy model πθ, search engine R, maximum action budget B. Ensure: Final response y. 1: Initialize rollout sequence y ←∅ 2: Initialize action count b ←0 3: while b < B do Initialize current action LLM rollout sequence yb ←∅ whi...
End of preview. Expand in Data Studio

No dataset card yet

Downloads last month
552