datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
FullStackBenchFullStack Bench: Evaluating LLMs as Full Stack Coders
Official repository for our paper "FullStack Bench: Evaluating LLMs as Full Stack Coders"
🏠 FullStack Bench Code •
📊 Benchmark Data •
📚 SandboxFusion
📌Introduction
FullStack Bench is a multilingual benchmark for full-stack programming, covering a wide range of application domains and 16 programming languages with 3K test samples, which substantially pushes… See the full description on the dataset page: https://huggingface.co/datasets/ByteDance/FullStackBench.fullstackbench-tasks
fullstackbench-tasks
Harbor task definitions used to benchmark coding agents in the FullStackBench / Loki research program. Each top-level directory is one self-contained task — Dockerfile, golden app, verifier, instruction — runnable via harbor run.
Prerequisites
hf CLI logged in to an ApolloTeam-member account (hf auth login)
Working install of Harbor — provides the harbor CLI (ask the Loki team if you don't already have it set up)
Docker daemon running locally
~4 GB… See the full description on the dataset page: https://huggingface.co/datasets/ApolloTeam/fullstackbench-tasks.fullstackbench-trajectories
fullstackbench-trajectories
FullStackBench tasks and the agent trajectories that ran them. One repo, two sibling trees keyed by the same task ID.
Layout
tasks/
└── <task-id>/ # task definition
├── environment/ # docker-compose, source app, fixtures
├── tests/ # verifier scripts
└── solution/ # reference impl (if applicable)
trajectories/
└── <task-id>/… See the full description on the dataset page: https://huggingface.co/datasets/ApolloTeam/fullstackbench-trajectories.FullStack-Bench
FullStack-Agent
Paper | Code | Dataset
Overview
This repository contains the FullStack-Bench dataset, introduced in the paper "FullStack-Agent: Enhancing Agentic Full-Stack Web Coding via Development-Oriented Testing and Repository Back-Translation".
In this paper, we propose FullStack-Agent, a unified system that combines a multi-agent full-stack development framework equipped with efficient coding and debugging tools (FullStack-Dev), an iterative self-improvement method… See the full description on the dataset page: https://huggingface.co/datasets/luzimu/FullStack-Bench.fullstackarena-gpt56sol-representatives-20260829
FullStackArena representative trajectories
This bundle contains one pass@1 model trajectory for each of 30 RideApp and 31
Arenagram template families, plus two explicitly labelled post-fix Arenagram
validation reruns. It does not claim that the model executed all 120 RideApp
or all 93 Arenagram instances.
Site
Phase
Raw evaluator
Adjudicated
rideapp
original_representative_run
25/30
28/30
arenagram
original_representative_run
22/31
22/31
arenagram… See the full description on the dataset page: https://huggingface.co/datasets/GroupieSteven/fullstackarena-gpt56sol-representatives-20260829.stargate_s04e01_100topkdiverse_text2vid
advanced-fullstack-ai-knowledge-base
Advanced Full-Stack & AI Engineering Knowledge Base (2026 Edition)
This repository contains a high-quality, production-ready sample subset of 23,734 records from a massive, proprietary dataset meticulously curated for Retrieval-Augmented Generation (RAG) systems, Agentic Workflows, and Fine-Tuning next-generation LLMs.
Overview & The Knowledge Cutoff Solution
One of the most persistent bottlenecks in production AI systems is the knowledge cutoff. Most… See the full description on the dataset page: https://huggingface.co/datasets/kooda-ai/advanced-fullstack-ai-knowledge-base.Robotwin-demo-clean-50-artifacts
RoboTwin beat_block_hammer demo_clean-50 Artifacts
这个仓库存放的是一次 RoboTwin 单任务闭环实验的关键结果产物，而不是官方完整数据集。
仓库内容
same_distribution/
demo_clean -> demo_clean 的评测结果与 100 个 rollout 视频
cross_configuration/
demo_clean -> demo_randomized 的评测结果与 100 个 rollout 视频
experiment_summary.json
已完成的闭环
本实验完整跑通了以下链路：
RoboTwin 官方任务程序生成专家轨迹
原始轨迹落盘为 raw data
ACT 预处理生成 training-ready data
ACT 正式训练
策略重新部署回 RoboTwin 环境
在 clean / randomized 两种配置下正式评测
关键结果
同分布：demo_clean ->… See the full description on the dataset page: https://huggingface.co/datasets/Full-Stack-Entity/Robotwin-demo-clean-50-artifacts.fullstackarena-gpt56sol-one-per-template-20260905
FullStackArena GPT-5.6-sol representative results
This release contains one immutable pass@1 trajectory for each maintained
template family: 30 RideApp trajectories and 31 Arenagram trajectories. It is
not a claim that all 120 RideApp or all
124 Arenagram instances in the two recorded
source packs were executed.
Website
Templates covered
Passed
Failed
Pass rate
RideApp
30/30
18
12
60.0%
Arenagram
31/31
17
14
54.8%
Outcome policy
Grades are… See the full description on the dataset page: https://huggingface.co/datasets/GroupieSteven/fullstackarena-gpt56sol-one-per-template-20260905.so101-left-sota-pack
SO101 Pickup-and-Putdown Left-Side SOTA Pack
项目简介
这个仓库汇总了当前阶段在 SO101 真机 pickup-and-putdown 任务上的三条成功策略路线，只保留左侧分区的阶段性最佳模型：
ACT
Diffusion Policy
SmolVLA
目标不是发布一个单一最佳 checkpoint，而是展示一套完整的具身实验工程成果：
数据采集
分区训练
真机闭环部署
rollout 录像
调试与复盘
当前这些成果已经足以说明：在同一任务、同一机器人平台、同一数据协议下，不同策略路线的行为特征与适配难点是可以被系统性比较与复现的。
任务定义
任务为：
从桌面抓取目标物体
将物体搬运到固定盒子上方
release 物体
完成放置
当前上传内容仅对应 left 分区，也就是目标物位于左侧工作区的场景。
与数据仓库的联动关系
本仓库只保存：
成功模型
rollout 视频
复盘文档
训练数据本体放在独立的数据仓库中：… See the full description on the dataset page: https://huggingface.co/datasets/Full-Stack-Entity/so101-left-sota-pack.fullstackarena-gpt56sol-one-per-template-20260902
FullStackArena GPT-5.6-sol representative results
This release contains one immutable pass@1 trajectory for each maintained
template family: 30 RideApp trajectories and 31 Arenagram trajectories. It is
not a claim that all 108 RideApp or all 116 Arenagram instances were executed.
Website
Templates covered
Passed
Failed
Pass rate
RideApp
30/30
18
12
60.0%
Arenagram
31/31
18
13
58.1%
Outcome policy
Grades are re-derived from every authoritative… See the full description on the dataset page: https://huggingface.co/datasets/GroupieSteven/fullstackarena-gpt56sol-one-per-template-20260902.synth-illuminati-cardsstage22so101-pickup-putdown-60fullstack-templatescik_sec_api_jsonlfull-stack-ai-engineering-datafullstack-templates-formattedfull_stack_dorm_100This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "so_follower",
"total_episodes": 44,
"total_frames": 12857,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 30,
"splits": {
"train": "0:44"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/H2Ozone/full_stack_dorm_100.stage3fullstackarena-gpt56sol-cross12-v700-20260910
FullStackArena GPT-5.6-sol cross-website v7 evaluation
This gated dataset contains the complete, credential-free evidence for one
pass over all 12 FullStackArena cross-website v7 tasks with
openai/gpt-5.6-sol through OpenRouter. The run used uncapped model waits,
the include-time protocol, a 900-second task timeout, one worker, and one
attempt per task.
Result
Metric
Value
Tasks
12
Passed
5
Failed
7
Infrastructure errors
0
Pass rate
41.7%… See the full description on the dataset page: https://huggingface.co/datasets/GroupieSteven/fullstackarena-gpt56sol-cross12-v700-20260910.fullstackarena-gpt56sol-cross-results-through-v8-20260916
FullStackArena cross-website results through v8
This gated dataset preserves the complete browser traces, screenshots, model-call records, evaluator outputs, protected state-check summaries, and integrity manifests for every cross-website GPT-5.6-sol task executed in the three reviewed releases below.
Results
Run group
Task-pack generation
Executed tasks
Result
Authoritative pass rate
Status
v7-cross12
Cross-site v7
12
5 pass / 7 fail
41.7%
All traces… See the full description on the dataset page: https://huggingface.co/datasets/GroupieSteven/fullstackarena-gpt56sol-cross-results-through-v8-20260916.fullstack-claim2-proxyfullstackarena-gpt56sol-cross4-refresh-v432-20260909
FullStackArena cross-site refresh qualification
This gated dataset contains the complete browser evidence for the four
ride_vs_active_mode cross-site instances affected by the RideApp refresh
contract correction. It is a focused qualification, not a new run of all 12
cross-site instances and not a full-pack pass-rate claim.
Immutable configuration
Model: openai/gpt-5.6-sol through OpenRouter.
FullStackArena source: 0910cc2085b1872238dd5f302ef3305035e309a0
from a… See the full description on the dataset page: https://huggingface.co/datasets/GroupieSteven/fullstackarena-gpt56sol-cross4-refresh-v432-20260909.full_stack_dormThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "so_follower",
"total_episodes": 103,
"total_frames": 33360,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 30,
"splits": {
"train": "0:103"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/H2Ozone/full_stack_dorm.fullstackarena-gpt56sol-rideapp-selective-v812-20260912
RideApp GPT-5.6-sol selective difficulty calibration
This gated evidence release contains the final observation selected for one
representative of each of RideApp's 30 task templates. It is not a fresh
30-task run. Following the registered PI-directed protocol, the 18 templates
that passed the September 5 baseline were rerun after difficulty refinement;
the 12 baseline failures were carried forward without another model call.
Three policy-sensitive observations and two… See the full description on the dataset page: https://huggingface.co/datasets/GroupieSteven/fullstackarena-gpt56sol-rideapp-selective-v812-20260912.so101-grasp-2so101-grasp-30fullstackarena-gpt56sol-auction25-five-replica-20260907
FullStackArena Auction: GPT-5.6-sol, one instance per template
This private evidence release contains the complete September 7, 2026 Auction
qualification run: one deterministic task from each of all 25 released
templates, executed with five isolated website/database/clock replicas.
Result interpretation
Raw framework result: 18/25 pass (72.0%).
Manual trace-audited result: 17/25 pass (68.0%).
Confirmed strict false positive: auction_454.
Confirmed strict false… See the full description on the dataset page: https://huggingface.co/datasets/GroupieSteven/fullstackarena-gpt56sol-auction25-five-replica-20260907.fullstackarena-gpt56sol-auction25-strict-v7-20260915
Auction GPT-5.6-sol strict-v7 selective difficulty calibration
This gated release represents all 25 Auction template families. It is not a
fresh 25-task run: gpt56sol-auction10-strict-v7-include-time-20260915-retry2 reruns only the ten strict-v6 representatives
that passed the independently audited baseline. Fifteen confirmed failures are
carried only after byte-level trace validation and exact representative plus
four-instance-family semantic equality across strict-v6 and… See the full description on the dataset page: https://huggingface.co/datasets/GroupieSteven/fullstackarena-gpt56sol-auction25-strict-v7-20260915.