The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
environment: struct<config: struct<checkpoint: string, t5_checkpoint: string, tokenizer: string, style_token: str (... 3065 chars omitted)
child 0, config: struct<checkpoint: string, t5_checkpoint: string, tokenizer: string, style_token: string, initialize (... 393 chars omitted)
child 0, checkpoint: string
child 1, t5_checkpoint: string
child 2, tokenizer: string
child 3, style_token: string
child 4, initializer: string
child 5, vectors: int64
child 6, lr: double
child 7, noise_shift: double
child 8, max_grad_norm: double
child 9, cpu_threads: int64
child 10, seed: int64
child 11, prompt: string
child 12, save_every: int64
child 13, checkpoint_policy: string
child 14, uncheckpointed_blocks: int64
child 15, fp32_text_context: bool
child 16, fused_rmsnorm: bool
child 17, fused_rope: bool
child 18, experiment: string
child 19, appearance_enabled: bool
child 20, lora_rank: int64
child 21, lora_alpha: int64
child 22, lora_lr: double
child 23, lora_dropout: double
child 1, checkpoint: struct<step: int64, conditioning: struct<cond_dim: null, adapter_type: string, adapter_hidden_dim: n (... 1948 chars omitted)
child 0, step: int64
child 1, conditioning: struct<cond_dim: null, adapter_type: string, adapter_hidden_dim: null, interaction_init: string, ful (... 1750 chars omitted)
child 0, cond_dim: null
child
...
_no_grad: bool
ddp_replicas_identical: bool
mean_step_seconds: double
mean_wall_seconds: double
mean_forward_seconds: double
mean_backward_seconds: double
mean_clips_per_second: double
optimizer_step: int64
forward_seconds: double
wall_seconds: double
ranks: list<item: struct<loss: double, grad_norm: double, vector_grad_norms: list<item: double>, embedding_ (... 343 chars omitted)
child 0, item: struct<loss: double, grad_norm: double, vector_grad_norms: list<item: double>, embedding_update_norm (... 331 chars omitted)
child 0, loss: double
child 1, grad_norm: double
child 2, vector_grad_norms: list<item: double>
child 0, item: double
child 3, embedding_update_norm: double
child 4, lora_b_grad_norm: double
child 5, forward_seconds: double
child 6, backward_seconds: double
child 7, optimizer_and_checks_seconds: double
child 8, step_seconds: double
child 9, peak_gpu_gib: double
child 10, output_shape: list<item: int64>
child 0, item: int64
child 11, frozen_parameters_have_no_grad: bool
child 12, rank: int64
child 13, sample_indices: list<item: int64>
child 0, item: int64
child 14, data_seconds: double
child 15, wall_seconds: double
loss: double
backward_seconds: double
peak_gpu_gib: double
parameter_rank_max_abs_difference: double
step_seconds: double
clips_per_second: double
warmup: bool
global_batch_size: int64
gradient_rank_max_abs_difference: double
to
{'optimizer_step': Value('int64'), 'warmup': Value('bool'), 'loss': Value('float64'), 'global_batch_size': Value('int64'), 'ranks': List({'loss': Value('float64'), 'grad_norm': Value('float64'), 'vector_grad_norms': List(Value('float64')), 'embedding_update_norm': Value('float64'), 'lora_b_grad_norm': Value('float64'), 'forward_seconds': Value('float64'), 'backward_seconds': Value('float64'), 'optimizer_and_checks_seconds': Value('float64'), 'step_seconds': Value('float64'), 'peak_gpu_gib': Value('float64'), 'output_shape': List(Value('int64')), 'frozen_parameters_have_no_grad': Value('bool'), 'rank': Value('int64'), 'sample_indices': List(Value('int64')), 'data_seconds': Value('float64'), 'wall_seconds': Value('float64')}), 'forward_seconds': Value('float64'), 'backward_seconds': Value('float64'), 'step_seconds': Value('float64'), 'wall_seconds': Value('float64'), 'peak_gpu_gib': Value('float64'), 'parameter_rank_max_abs_difference': Value('float64'), 'gradient_rank_max_abs_difference': Value('float64'), 'clips_per_second': Value('float64')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
environment: struct<config: struct<checkpoint: string, t5_checkpoint: string, tokenizer: string, style_token: str (... 3065 chars omitted)
child 0, config: struct<checkpoint: string, t5_checkpoint: string, tokenizer: string, style_token: string, initialize (... 393 chars omitted)
child 0, checkpoint: string
child 1, t5_checkpoint: string
child 2, tokenizer: string
child 3, style_token: string
child 4, initializer: string
child 5, vectors: int64
child 6, lr: double
child 7, noise_shift: double
child 8, max_grad_norm: double
child 9, cpu_threads: int64
child 10, seed: int64
child 11, prompt: string
child 12, save_every: int64
child 13, checkpoint_policy: string
child 14, uncheckpointed_blocks: int64
child 15, fp32_text_context: bool
child 16, fused_rmsnorm: bool
child 17, fused_rope: bool
child 18, experiment: string
child 19, appearance_enabled: bool
child 20, lora_rank: int64
child 21, lora_alpha: int64
child 22, lora_lr: double
child 23, lora_dropout: double
child 1, checkpoint: struct<step: int64, conditioning: struct<cond_dim: null, adapter_type: string, adapter_hidden_dim: n (... 1948 chars omitted)
child 0, step: int64
child 1, conditioning: struct<cond_dim: null, adapter_type: string, adapter_hidden_dim: null, interaction_init: string, ful (... 1750 chars omitted)
child 0, cond_dim: null
child
...
_no_grad: bool
ddp_replicas_identical: bool
mean_step_seconds: double
mean_wall_seconds: double
mean_forward_seconds: double
mean_backward_seconds: double
mean_clips_per_second: double
optimizer_step: int64
forward_seconds: double
wall_seconds: double
ranks: list<item: struct<loss: double, grad_norm: double, vector_grad_norms: list<item: double>, embedding_ (... 343 chars omitted)
child 0, item: struct<loss: double, grad_norm: double, vector_grad_norms: list<item: double>, embedding_update_norm (... 331 chars omitted)
child 0, loss: double
child 1, grad_norm: double
child 2, vector_grad_norms: list<item: double>
child 0, item: double
child 3, embedding_update_norm: double
child 4, lora_b_grad_norm: double
child 5, forward_seconds: double
child 6, backward_seconds: double
child 7, optimizer_and_checks_seconds: double
child 8, step_seconds: double
child 9, peak_gpu_gib: double
child 10, output_shape: list<item: int64>
child 0, item: int64
child 11, frozen_parameters_have_no_grad: bool
child 12, rank: int64
child 13, sample_indices: list<item: int64>
child 0, item: int64
child 14, data_seconds: double
child 15, wall_seconds: double
loss: double
backward_seconds: double
peak_gpu_gib: double
parameter_rank_max_abs_difference: double
step_seconds: double
clips_per_second: double
warmup: bool
global_batch_size: int64
gradient_rank_max_abs_difference: double
to
{'optimizer_step': Value('int64'), 'warmup': Value('bool'), 'loss': Value('float64'), 'global_batch_size': Value('int64'), 'ranks': List({'loss': Value('float64'), 'grad_norm': Value('float64'), 'vector_grad_norms': List(Value('float64')), 'embedding_update_norm': Value('float64'), 'lora_b_grad_norm': Value('float64'), 'forward_seconds': Value('float64'), 'backward_seconds': Value('float64'), 'optimizer_and_checks_seconds': Value('float64'), 'step_seconds': Value('float64'), 'peak_gpu_gib': Value('float64'), 'output_shape': List(Value('int64')), 'frozen_parameters_have_no_grad': Value('bool'), 'rank': Value('int64'), 'sample_indices': List(Value('int64')), 'data_seconds': Value('float64'), 'wall_seconds': Value('float64')}), 'forward_seconds': Value('float64'), 'backward_seconds': Value('float64'), 'step_seconds': Value('float64'), 'wall_seconds': Value('float64'), 'peak_gpu_gib': Value('float64'), 'parameter_rank_max_abs_difference': Value('float64'), 'gradient_rank_max_abs_difference': Value('float64'), 'clips_per_second': Value('float64')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
CS:GO 风格数据:最低 v2 协议
所有正式风格采集必须至少满足 external_v2 完整数据协议;只有 RGB 视频或 RGB/VAE 缓存不予收录。
白模视频是额外产物,不替代 RGB、Dense、状态、交互或相机真值;需要覆盖环境、第一人称手和武器,
从真实引擎渲染,并与 RGB 使用相同 replay、ego、tick、相机与帧时间。
原生 GT / 白模采集批次
统一总目录:original/batches/。各批次独立命名,保留旧路径;时长按配对时间线计一次。
| 批次 | 已发布 case | 配对时长 | 片段长度 | 状态 |
|---|---|---|---|---|
| 20261004:原有 AK v2 白模批次 | 282 | 23.50 分钟 | 5 秒切片 | 以该批 verified_manifest 为准 |
| 20261005:六客户端常驻批次 | 192 | 64.00 分钟 | 20 秒 | 采集中,每 30 分钟增量同步 |
新批次包含 GT、白模视频和 external_v2 完整字段;321 个 RGB/白模帧、81 个 Dense 帧,832×480、16 FPS。
仅将已完成本地验证且通过 HF 固定提交 SHA256 全量读回的 case 纳入发布清单。失败候选不上传,白模不重复计时。
抽样画面已获用户确认;全批视觉、像素一致性和碰撞验收尚未完成,保留 training_ready=false。
原有审计中“风格有效数据为 0”的结论指当时的 MOD 风格数据,不代表本节原生采集归档为空。
更新于 2026-10-05T17:21:30.167042+00:00。详细 SHA、来源与分组见各批次 verified_manifest.json。
梵高星月夜风格训练配对批次(2026-10-05)
本批包含 66 个独立5秒片段,共330秒(5分30秒):24条自身参考重做、41条后续生成、1条历史试验。仅代表仓库的一个子集。66条均找到已有HF原始v2源包,指纹一致,缺少配对0条。已有原始v2、原画与白模通过固定版本和内容指纹复用,不重新上传;本批仅新增66条832×480风格MP4(约241MB)和配对/时间/生成/核验索引。 训练数据由已有完整external_v2条件包与新增风格目标共同组成。生成原件与参考过程素材保留本地,不重复上传为训练附件。
原始RGB/白模81帧16fps,Dense/State/Interaction为21个时间点;风格目标150帧30fps,需要按名义时间显式适配训练入口,不能按帧号覆盖原GT。66条原始必需字段和声明时间索引通过核验;51条另有逐帧引擎相机证据,15条早期源包无这项额外归档。生成像素/几何对齐未认证,保留派生训练候选及严格v2未认证标记。24条有用户视觉通过记录,41条待视觉审核,1条历史输入证据不完整。发布逐文件全量读回收据见批次 publication-verification.json。
当前有效风格数据
2026-10-03 审计:本仓库有效 v2 风格数据 0 cases / 0 小时。
旧 Halloween 360 条 × 10 秒(1 小时)只有 RGB,未包含必需的 Dense/State/Interaction/Camera,
两批 csgo_rgb_vae_cache_v1 也明确缺少条件字段。按所有者要求已从当前分支删除旧
halloween/ 与 processed/halloween/ 下的不完整数据及缓存。
删除清单、固定源修订和缺失证据在 audits/style_v2_cleanup_20261003.json。
删除为 Git 版本提交:历史修订仍可恢复;不是抹除历史或保证释放历史存储额度。
Cyberpunk 和其他 MOD 目前没有本仓库正式数据,本地审核预览不计入有效数量。
必需产物与验收
每个训练窗口必须有真实、帧对齐的 window00/gt.mp4、dense/*.npz、
dense_manifest.jsonl、player_mask_config.json、state/state_interaction.npz、
interaction.npz、poses.npy、intrinsics.npy、prompt.txt。
Dense 使用 full7_yawangle_v2、有限 float32 [7,240,416];早期 5 秒窗口为 81 RGB 帧、21 Dense/条件帧;本次 20 秒窗口为 321 RGB 帧、81 Dense/条件帧。
具体以批次 manifest 为准,依据实际协议采样,不能补造或重复尾帧。
使用 real-player-mask,记录实际隐藏玩家策略;隐藏玩家不等于省略状态或 Dense 文件。
必须校验 RGB/Dense/State/Interaction/Camera 的身份、tick、shape、有限值和帧对齐。
预处理缓存必须含 latent、i2v_y、dense_features、state、interaction、
camera_poses、camera_intrinsics、text_emb,并保留精确 prompt/T5、模型/源码 SHA、
源 revision/path/样本 ID、分片及样本 SHA、split、校验收据;RGB-only 缓存禁止作为正式产物。
采集和渲染配置需不可变 Git/SHA 跟踪,B 包与 B 门附近范围逐 tick 验收,保留武器手并隐藏其他玩家。
QA、解码与完整 SHA 读回通过才计入有效数据;坏 case 不凑数、不降低阈值。
后续路径约定(尚未发布,不表示已有数据)
<style>/batches/<immutable_version>/shards/:完整 v2 case;
<style>/batches/<immutable_version>/manifest.json:实际验证样本与源配置;
extras/white_model/:按 case ID 对齐的白模视频与对应 manifest;
processed/<style>/<immutable_version>/full_conditions_v2/:完整条件预处理缓存。
白模应随所属完整 case 保存和校验,不能单独算额外采集时长。
保留的原生数据与缓存
本次只删除上述确认缺字段的风格数据;processed/native/ 完整条件缓存以及
三个原始 faraway00/csgo-training-cases-part-{1,2,3} 仓库未修改。
私有武器切换仓库未修改可见性或删除。原生缓存的实际完成数以各目录 manifest 为准;
以下为既有流程说明,不是本次全量合格认证。
完整条件:csgo_training_cache_features_v1
完整原始 case 包含 window00/gt.mp4、dense/*.npz、dense_manifest.jsonl、
state/state_interaction.npz、interaction.npz、poses.npy、intrinsics.npy、prompt.txt。
仅接受 real-player-mask 与 full7_yawangle_v2,严格校验 Dense/State/Interaction/RGB/Camera 的帧对齐。
未知 prompt 必须补充真实 T5 编码;不能用其他 prompt 的 embedding 替换。
输出保留 latent、i2v_y、dense_features、state、interaction、camera_poses、camera_intrinsics、text_emb;
不保留原始 Dense7。前七项 FP32,text_emb BF16。
完整条件索引同时提供 manifest_schema=csgo_cache_manifest_v1、固定分片 revision、
分片与每个 .pt 的 SHA256/字节数、样本顺序以及 split 字段,可由对应 HF TAR reader 消费。
当前转换索引全部标为 train,尚未创建独立验证集;这不是随机 train/val 切分。
使用训练器时须显式加载对应子目录的 manifest,不能将本仓库根目录当作单一训练缓存。
批处理中已发现部分源 case 缺少 player_mask_config.json,不能满足真实玩家 mask 校验。
不兼容输入记录在 manifest 的 rejected_samples,保留源路径与原因,不补造字段、不计入有效样本。
因此 3,030 是原始 case 总数,而不是保证可用的完整条件缓存数量。
原始完整 case 标签:fire_reload_no_switch 1,018、weapon_switch 500、
camera_end_changed 881、camera_turned_returned 315、no_interaction 500。
标签存在交叉,只有 no_interaction 与其他标签不重合;不能直接加总为 case 数。
源清单路径为各分仓 manifest/cases_this_part.jsonl,总清单 manifest/cases_all.jsonl。
5. 流式读取与可追溯性
适配器通过 HF 固定 revision 的 HTTP stream=True 顺序读取,不做全仓 snapshot 下载:
RGB 一次中转一条 MP4,完整条件 TAR 一次还原一个 case;处理后清理临时输入。
每个输出样本记录来源 repo/revision/path、帧范围、样本 ID、shape/dtype、模型 SHA 与设备型号。
TAR 分片上传后做远端大小/LFS SHA 核对及完整 HTTP 字节读回,张量 roundtrip/有限值验证通过才纳入 manifest。
上传失败时保留待上传产物,不删除尚未验证的数据。
HTTP 连接断开时按字节 Range 续读;任务异常时从已验证 manifest 恢复,避免重复发布已完成样本。
当前批任务分配为 GPU 1 处理 Halloween,GPU 0 顺序处理原始完整条件 part 1→2→3;
此说明记录启动配置,并非实时运行状态。最终完成须见 manifest 的 status=complete;
若为 complete_with_rejections,仅表示扫描完毕,仍须处理拒绝清单中的源数据问题。
源码在 Pod 网络盘 /workspace/state2view-hf-cache-pipeline;
流式适配器在 /workspace/state2view-hf-preprocess/stream_hf_preprocess.py;
中转、环境与日志在根盘 /opt/state2view-hf-preprocess*。
网络盘不支持 chmod,Git 元数据位于 /opt/state2view-hf-cache-pipeline-git;重建 Pod 后需恢复元数据。
权利说明
数据来自 CS:GO 及第三方地图 MOD,不授予游戏或地图资产的额外版权许可。
Halloween 原生采集批次(2026-10-05,非完整 V2)
新增 单人物随机跟随30分钟批次:90段,每段20秒,800×450、16FPS。保留 Halloween 地图风格;可见人物、第一人称手和武器为白模。人物在较宽路线实际行走,客户端真实移动跟随,并平滑随机改变左右构图和正面/侧面/背面朝向。
本批提供白模视频、原生人物实例掩码、深度、实体骨架、相机诊断、源 demo、逐 tick 状态/动作/事件和验收收据,属于原生回放配对采集证据。full_v2_complete=false,源快照相位未认证,缺少完整协议 Dense/State/Interaction/Camera 和原始 RGB 窗口;不纳入上方完整 V2 正式数据 case/小时统计。完整 V2 训练要求继续适用。
视频可直接浏览,各段原生数据与对应源资料集中在6个 TAR 分片中;manifest 记录分片/视频/配置哈希与逐段验证,逐段索引 记录原生分片路径。仅发布最新通过验收的30分钟,早期试采不计入。
Halloween 完整 V2 采集输入补齐(2026-10-05)
本批说明 · 完整索引:远端完整SHA256读回已验证 90段、1800秒,每段20秒,321 RGB/相机帧、81 Dense/State/Interaction端点,附严格对齐白模。状态 complete;目标90段30分钟。
V2采集输入字段齐全,未生成模型缓存,作者像素/在线快照相位认证未完成,因此 training_ready=false,不改变上方正式训练资格统计。原raw批次仍保留;新批次采用Halloween实际BSP与原生NAV,原始RGB与白模相机/第一人称骨骼逐帧配对。
- Downloads last month
- 2,122