step0-all
收藏资源简介:
本数据集名为step0-all,是专门为ULVR_v2_clean数据集生成的预计算教师潜在表示(步骤0),核心目的是为Monet模型的监督微调(SFT)提供知识蒸馏目标。数据来源于ULVR_v2_clean数据集的四个子类别:text_cot、helper_interleaved、scene_graph和segmentation(排除了depth、edge、bbox_crop、bbox_highlight子集),每个子类别均包含训练集和验证集。生成过程使用了经过Monet补丁的Qwen2.5-VL-7B-Instruct基础模型,在潜在模式下对每个训练样本进行前向传播,记录每个辅助图像后跟随的潜在槽(latent_size = 8)在所有29个Transformer层中的隐藏状态。数据集以Parquet格式组织,主要字段包括:标识源样本的sample_id字符串、表示所属类别的category字符串、指明原始数据集的source_dataset字符串、表示辅助图像数量的num_intermediate_steps整数、原始张量数据类型latent_dtype(如bfloat16)、描述潜在向量维度的latent_shape列表(格式为[层数, 潜在令牌数, 隐藏维度],例如[29, 8, 3584]),以及扁平化存储的潜在向量latent列表。该数据集主要用于多模态视觉推理、视觉问答和潜在表示学习等任务,作为训练过程中模型对齐和知识传递的监督信号。
The dataset named step0-all is a precomputed teacher latent representation (step 0) specifically generated for the ULVR_v2_clean dataset. Its core purpose is to provide knowledge distillation targets for the supervised fine-tuning (SFT) of the Monet model. The data is sourced from four subcategories of the ULVR_v2_clean dataset: text_cot, helper_interleaved, scene_graph, and segmentation (excluding the depth, edge, bbox_crop, and bbox_highlight subsets), each containing training and validation sets. The generation process uses the Monet-patched Qwen2.5-VL-7B-Instruct base model in latent mode to perform forward propagation on each training sample. During this process, each auxiliary (intermediate) image is followed by `latent_size = 8` latent slots, and the hidden states at these latent positions across all 29 Transformer layers are recorded. The dataset is organized in Parquet format, with key fields including: a `sample_id` string identifying the source sample, a `category` string indicating the category, a `source_dataset` string specifying the original dataset, an integer `num_intermediate_steps` for the number of auxiliary images, the original tensor data type `latent_dtype` (e.g., bfloat16), a `latent_shape` list describing the latent vector dimensions (in the format [number of layers, number of latent tokens, hidden dimension], e.g., [29, 8, 3584]), and a flattened list `latent` storing the latent vectors. This dataset is primarily used for tasks such as multimodal visual reasoning, visual question answering, and latent representation learning, serving as a supervision signal for model alignment and knowledge transfer during training.
数据集概述
step0-all 是一个预先计算好的 教师潜变量(teacher latents) 数据集,专为视觉推理和多模态学习任务设计。
基本信息
- 许可证: Apache-2.0
- 任务类别: 视觉问答(visual-question-answering)
- 标签: 视觉推理、多模态、潜变量推理、教师潜变量
- 来源数据集: RuoliuYang/ULVR_v2_clean(排除了 depth、edge、bbox_crop、bbox_highlight 子集)
- 教师模型: Qwen2.5-VL-7B-Instruct(Monet 补丁版本)
数据集构成
该数据集包含 4 个配置(config),每个配置均有训练集和验证集:
| 配置名称 | 训练集路径 | 验证集路径 |
|---|---|---|
| text_cot | data/text_cot/*.parquet | data/text_cot_val/*.parquet |
| helper_interleaved | data/helper_interleaved/*.parquet | data/helper_interleaved_val/*.parquet |
| scene_graph | data/scene_graph/*.parquet | data/scene_graph_val/*.parquet |
| segmentation | data/segmentation/*.parquet | data/segmentation_val/*.parquet |
数据生成原理
对于每个训练样本,教师模型在潜变量模式下执行前向传播:
- 每个辅助(中间)图像后跟随
latent_size = 8个潜变量槽位 - 记录这些潜变量位置在 全部 29 层 中的隐藏状态
- 这些潜变量可作为 Monet SFT 的蒸馏目标
数据列说明
| 列名 | 类型 | 含义 |
|---|---|---|
| sample_id | string | 匹配原始样本的 ID |
| category | string | 类别之一(共 4 类) |
| source_dataset | string | 原始来源数据集名称 |
| num_intermediate_steps | int64 | 辅助图像的数量 |
| latent_dtype | string | 原始张量数据类型(bfloat16) |
| latent_shape | list<int> | 潜变量形状,例如 [29, 8, 3584] 或 [29, 8*N, 3584](N 为辅助图像数) |
| latent | list<float16> | 展平的潜变量,需根据 latent_shape 重塑还原 |
使用示例
python import numpy as np from datasets import load_dataset
ds = load_dataset("RuoliuYang/step0-all", "scene_graph", split="train", streaming=True) row = next(iter(ds)) lat = np.array(row["latent"], dtype=np.float16).reshape(row["latent_shape"])




