You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

CoReBench 原始 Rollout

这里包含 DeepSeek 在同一批1080个 CoReBench 任务上的两轮图像生成实验:一轮调用 Z-Image,另一轮调用 Qwen-Image-2.1。可以查看每个任务的完整思考与工具调用过程、每轮生成图片、验收反馈和最终选择;同时附有两轮对应的实验代码。

实验 原始轨迹 逐轮图片 agent 自检通过
DeepSeek + Z-Image 1080 3216 562
DeepSeek + Qwen-Image-2.1 1080 2207 939

Actor 均为 DeepSeek 4.1 Flash,thinking=high,seed=0,每个任务最多生成4次。自检通过数不是独立 benchmark 分数。所有结果均保留,包括预算耗尽、未验收和未完成的任务。

从哪里开始看

两轮数据分别位于:

  • Z-Image:T2I-CoReBench/20260923-full1080-deepseek41flash-zimage-conc6/
  • Qwen-Image-2.1:T2I-CoReBench/20260928-150500-full1080-deepseek41flash-qwen21-conc6/

进入其中一轮,先打开 cases.portable.jsonl。每行对应一个任务,包含任务 ID、原始 prompt、轨迹路径、每轮图片路径和最终选图路径。所有路径都相对于这一轮的目录。

想看什么 文件或目录
原始任务及图片位置 cases.portable.jsonl
完整轨迹 output/cases/T2I-CoReBench/<split>/<ID>/session.jsonl
第1轮、第2轮等生成图片 同一案例目录的 round-01/image.jpg、round-02/image.jpg 等
每轮生成参数和结果 对应 round-XX/result.json
案例最终结果 案例目录的 result.json
全部任务的原始运行结果 output/results.jsonl
实验代码 code/

session.jsonl 保留原始事件,包含用户输入、assistant 的 thinking 和文本、工具调用参数、工具返回、图片及工作流状态。它不是纯聊天文本;恢复对话分支时使用事件的 parentId。结合逐轮图片,可以查看 agent 如何生成、判断和修订。

读取一个案例

在 Raw_Rollouts/ 下执行:

import json
from pathlib import Path

run = Path('T2I-CoReBench/20260928-150500-full1080-deepseek41flash-qwen21-conc6')
with (run / 'cases.portable.jsonl').open() as f:
    case = json.loads(next(f))

print('任务:', case['id'])
print('Prompt:', case['prompt'])
print('轨迹:', run / case['session'])
for i, image in enumerate(case['round_images'], 1):
    print(f'第{i}轮图片:', run / image)
if case['selected_image']:
    print('最终选图:', run / case['selected_image'])

原始 session 和 result 中仍有原机器的绝对路径;在自己的机器上使用 cases.portable.jsonl 的相对路径即可。查看数据无需安装 pi 或启动模型服务。

查看实验实现

每轮 code/ 中,优先看以下文件:

  • .pi/SYSTEM.md:agent 的行为说明。
  • .pi/extensions/image-gen.ts:调用生图后端。
  • .pi/extensions/image-workflow.ts:清单、验收和最终选择。
  • scripts/batch-agent-image.mjs:批量任务执行。

本目录仅包含原始 rollout、生图产物和实验代码,不包含训练数据或模型权重。两轮文件合计约4.93 GB。交付仓库:https://huggingface.co/datasets/wxli318/GenAgent_Raw_Rollouts 。

Downloads last month
-