The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
video_id: string
prompt: string
seed: int64
source: string
width: int64
height: int64
num_frames: int64
reward_backend: string
lora_rank: int64
lora_alpha: int64
lora_blocks: string
exclude_modules: null
peft_version: string
loftq_config: struct<>
corda_config: null
trainable_token_indices: null
auto_mapping: struct<base_model_class: string, parent_library: string>
child 0, base_model_class: string
child 1, parent_library: string
bias: string
megatron_config: null
arrow_config: null
peft_type: string
ensure_weight_tying: bool
megatron_core: string
alpha_pattern: struct<>
lora_dropout: double
use_dora: bool
inference_mode: bool
layer_replication: null
target_parameters: null
modules_to_save: null
base_model_name_or_path: null
fan_in_fan_out: bool
revision: null
eva_config: null
init_lora_weights: bool
r: int64
task_type: null
layers_pattern: null
alora_invocation_tokens: null
rank_pattern: struct<>
use_rslora: bool
lora_bias: bool
qalora_group_size: int64
target_modules: list<item: string>
child 0, item: string
layers_to_transform: null
use_qalora: bool
to
{'alora_invocation_tokens': Value('null'), 'alpha_pattern': {}, 'arrow_config': Value('null'), 'auto_mapping': {'base_model_class': Value('string'), 'parent_library': Value('string')}, 'base_model_name_or_path': Value('null'), 'bias': Value('string'), 'corda_config': Value('null'), 'ensure_weight_tying': Value('bool'), 'eva_config': Value('null'), 'exclude_modules': Value('null'), 'fan_in_fan_out': Value('bool'), 'inference_mode': Value('bool'), 'init_lora_weights': Value('bool'), 'layer_replication': Value('null'), 'layers_pattern': Value('null'), 'layers_to_transform': Value('null'), 'loftq_config': {}, 'lora_alpha': Value('int64'), 'lora_bias': Value('bool'), 'lora_dropout': Value('float64'), 'megatron_config': Value('null'), 'megatron_core': Value('string'), 'modules_to_save': Value('null'), 'peft_type': Value('string'), 'peft_version': Value('string'), 'qalora_group_size': Value('int64'), 'r': Value('int64'), 'rank_pattern': {}, 'revision': Value('null'), 'target_modules': List(Value('string')), 'target_parameters': Value('null'), 'task_type': Value('null'), 'trainable_token_indices': Value('null'), 'use_dora': Value('bool'), 'use_qalora': Value('bool'), 'use_rslora': Value('bool')}
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 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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
video_id: string
prompt: string
seed: int64
source: string
width: int64
height: int64
num_frames: int64
reward_backend: string
lora_rank: int64
lora_alpha: int64
lora_blocks: string
exclude_modules: null
peft_version: string
loftq_config: struct<>
corda_config: null
trainable_token_indices: null
auto_mapping: struct<base_model_class: string, parent_library: string>
child 0, base_model_class: string
child 1, parent_library: string
bias: string
megatron_config: null
arrow_config: null
peft_type: string
ensure_weight_tying: bool
megatron_core: string
alpha_pattern: struct<>
lora_dropout: double
use_dora: bool
inference_mode: bool
layer_replication: null
target_parameters: null
modules_to_save: null
base_model_name_or_path: null
fan_in_fan_out: bool
revision: null
eva_config: null
init_lora_weights: bool
r: int64
task_type: null
layers_pattern: null
alora_invocation_tokens: null
rank_pattern: struct<>
use_rslora: bool
lora_bias: bool
qalora_group_size: int64
target_modules: list<item: string>
child 0, item: string
layers_to_transform: null
use_qalora: bool
to
{'alora_invocation_tokens': Value('null'), 'alpha_pattern': {}, 'arrow_config': Value('null'), 'auto_mapping': {'base_model_class': Value('string'), 'parent_library': Value('string')}, 'base_model_name_or_path': Value('null'), 'bias': Value('string'), 'corda_config': Value('null'), 'ensure_weight_tying': Value('bool'), 'eva_config': Value('null'), 'exclude_modules': Value('null'), 'fan_in_fan_out': Value('bool'), 'inference_mode': Value('bool'), 'init_lora_weights': Value('bool'), 'layer_replication': Value('null'), 'layers_pattern': Value('null'), 'layers_to_transform': Value('null'), 'loftq_config': {}, 'lora_alpha': Value('int64'), 'lora_bias': Value('bool'), 'lora_dropout': Value('float64'), 'megatron_config': Value('null'), 'megatron_core': Value('string'), 'modules_to_save': Value('null'), 'peft_type': Value('string'), 'peft_version': Value('string'), 'qalora_group_size': Value('int64'), 'r': Value('int64'), 'rank_pattern': {}, 'revision': Value('null'), 'target_modules': List(Value('string')), 'target_parameters': Value('null'), 'task_type': Value('null'), 'trainable_token_indices': Value('null'), 'use_dora': Value('bool'), 'use_qalora': Value('bool'), 'use_rslora': Value('bool')}
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.
LTX-Video-13B GRPO — 500-set PILOT (5 prompts)
Pilot for review before committing the full 500-prompt run. 5 prompts (FA_001–FA_005), GRPO only, no baseline arm.
Pairs with angelic123/ltx13b_baseline
(496 videos, the 500-set baseline arm). Seeds here are the original 500-set line numbers
(FA_001→2 … FA_005→6), matching the baseline's --base_seed = LINE_NUM, so these 5 are seed-paired
with their baseline counterparts.
Verified run configuration
Read from the executed torchrun command line, cross-checked against all 5 metadata.json and the
live reward payloads.
| parameter | value |
|---|---|
| model | Lightricks/LTX-Video, config ltxv-13b-0.9.8-dev.yaml |
| parallelism | 4× H100, Ulysses SP (--ltx-ulysses-size 4), one prompt at a time |
| diffusion (denoising) steps | 50 |
| GRPO window | last 15 timesteps (35..49) |
| rollouts per GRPO step | 5 — independent random seed + guidance scale ∈ [3.0, 12.0] each |
| reward | staged: first 5 GRPO steps image-CLIP, remaining 10 Qwen/Qwen2-VL-7B-Instruct (temp 0.15) |
| LoRA | rank 4, alpha 8, blocks last, unfreeze 0.20 |
| LR | 1e-4 |
| guidance (base) | 7.5 |
| output | 704×480, 32 frames, 8 fps, 4.0 s |
Total rollouts: 375 (5 prompts × 15 steps × 5), with 375 distinct seeds — verified independent.
⚠️ Geometry differs from the baseline
| resolution | frames | fps | |
|---|---|---|---|
| this pilot (GRPO) | 704×480 | 32 | 8 |
ltx13b_baseline |
720×480 | 32 | 8 |
LTX's VAE has a 32× spatial stride and 720 is not a multiple of 32 (720/32 = 22.5), so the pipeline snaps it to 704; LTX-13B cannot emit 720 natively. No resize was applied — rescaling only one arm would bias pixel metrics and the file would stop being what the model produced. For PSNR/SSIM/LPIPS against the baseline, resample both arms to a common size at eval time. VLM-scored or human comparison is unaffected.
Reward telemetry
Qwen-phase (GRPO steps 6–15) mean reward, first → last:
| prompt | first | last | std range |
|---|---|---|---|
| FA_001 | 0.6567 | 0.6027 | 0.0000–0.0392 |
| FA_002 | 0.6667 | 0.6567 | 0.0000–0.1158 |
| FA_003 | 0.6667 | 0.6667 | 0.0000–0.0200 |
| FA_004 | 0.6347 | 0.5753 | 0.0000–0.0293 |
| FA_005 | 0.7612 | 0.6167 | 0.0000–0.1166 |
These are NOT a learning curve. num_grpo_steps indexes trailing denoising timesteps, so each
step scores a video continued from a different noise level — step 6 and step 15 are not comparable.
Judging whether GRPO improved anything requires scoring the final videos against the baseline's with
the same teacher.
Advantage resolution is a real limitation: reward std reaches exactly 0.0000 on every prompt
at some step — all 5 rollouts scoring identically, giving zero advantage signal. Qwen pins
text_alignment and dynamic_motion_consistency at 10.0 and moves only physical_plausibility in
half-point increments, so a 5-rollout group frequently collapses to 1–2 distinct values. Sampling is
not the cause (375/375 distinct seeds), after fixing a bug where the RNG was re-seeded from
base_seed every step and advanced by only step_index, while each rollout consumes two draws —
that yielded just 15 distinct seeds across 75 rollouts with 59 (seed, guidance) pairs shared between
adjacent steps.
Contents
FA_00{1..5}.mp4— GRPO outputsmanifest.json— per video: prompt, seed, source path, geometry, LoRA configlora_adapters/FA_00{1..5}.mp4/— per-prompt trained LoRA (rank 4), the actual trained artifact
- Downloads last month
- 46