Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
version: int64
taskId: string
category: string
subtask: string
rolloutId: int64
model: string
provider: string
seed: int64
question: string
messages: list<item: struct<role: string, content: list<item: struct<type: string, text: string, id: string, n (... 83 chars omitted)
child 0, item: struct<role: string, content: list<item: struct<type: string, text: string, id: string, name: string (... 71 chars omitted)
child 0, role: string
child 1, content: list<item: struct<type: string, text: string, id: string, name: string, arguments: struct<path: stri (... 22 chars omitted)
child 0, item: struct<type: string, text: string, id: string, name: string, arguments: struct<path: string, command (... 10 chars omitted)
child 0, type: string
child 1, text: string
child 2, id: string
child 3, name: string
child 4, arguments: struct<path: string, command: string>
child 0, path: string
child 1, command: string
child 2, timestamp: int64
providerExposedReasoning: list<item: null>
child 0, item: null
events: list<item: struct<type: string, timestamp: int64, payload: struct<role: string, updateType: string, (... 349 chars omitted)
child 0, item: struct<type: string, timestamp: int64, payload: struct<role: string, updateType: string, stopReason: (... 337 chars omitted)
child 0, type: string
child 1, timestamp: int64
child 2, payload: struct<
...
nd: string
child 3, result: struct<content: list<item: struct<type: string, text: string, mimeType: string, file: string>>>
child 0, content: list<item: struct<type: string, text: string, mimeType: string, file: string>>
child 0, item: struct<type: string, text: string, mimeType: string, file: string>
child 0, type: string
child 1, text: string
child 2, mimeType: string
child 3, file: string
child 4, isError: bool
child 5, startedAt: int64
child 6, endedAt: int64
artifacts: list<item: struct<path: string, mediaType: string, sha256: string, sizeBytes: int64>>
child 0, item: struct<path: string, mediaType: string, sha256: string, sizeBytes: int64>
child 0, path: string
child 1, mediaType: string
child 2, sha256: string
child 3, sizeBytes: int64
answer: string
normalizedAnswer: string
reward: int64
exitReason: string
usage: struct<inputTokens: int64, outputTokens: int64>
child 0, inputTokens: int64
child 1, outputTokens: int64
runtime: struct<wallSeconds: double>
child 0, wallSeconds: double
perception: struct<sam31Calls: int64, cotracker3Calls: int64, da3Calls: int64, vggtOmegaCalls: int64, gpuSeconds (... 32 chars omitted)
child 0, sam31Calls: int64
child 1, cotracker3Calls: int64
child 2, da3Calls: int64
child 3, vggtOmegaCalls: int64
child 4, gpuSeconds: struct<>
child 5, cacheHits: struct<>
trajectoryPath: string
to
{'taskId': Value('string'), 'rolloutId': Value('int64'), 'answer': Value('string'), 'normalizedAnswer': Value('string'), 'reward': Value('int64'), 'exitReason': Value('string'), 'trajectoryPath': Value('string')}
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
version: int64
taskId: string
category: string
subtask: string
rolloutId: int64
model: string
provider: string
seed: int64
question: string
messages: list<item: struct<role: string, content: list<item: struct<type: string, text: string, id: string, n (... 83 chars omitted)
child 0, item: struct<role: string, content: list<item: struct<type: string, text: string, id: string, name: string (... 71 chars omitted)
child 0, role: string
child 1, content: list<item: struct<type: string, text: string, id: string, name: string, arguments: struct<path: stri (... 22 chars omitted)
child 0, item: struct<type: string, text: string, id: string, name: string, arguments: struct<path: string, command (... 10 chars omitted)
child 0, type: string
child 1, text: string
child 2, id: string
child 3, name: string
child 4, arguments: struct<path: string, command: string>
child 0, path: string
child 1, command: string
child 2, timestamp: int64
providerExposedReasoning: list<item: null>
child 0, item: null
events: list<item: struct<type: string, timestamp: int64, payload: struct<role: string, updateType: string, (... 349 chars omitted)
child 0, item: struct<type: string, timestamp: int64, payload: struct<role: string, updateType: string, stopReason: (... 337 chars omitted)
child 0, type: string
child 1, timestamp: int64
child 2, payload: struct<
...
nd: string
child 3, result: struct<content: list<item: struct<type: string, text: string, mimeType: string, file: string>>>
child 0, content: list<item: struct<type: string, text: string, mimeType: string, file: string>>
child 0, item: struct<type: string, text: string, mimeType: string, file: string>
child 0, type: string
child 1, text: string
child 2, mimeType: string
child 3, file: string
child 4, isError: bool
child 5, startedAt: int64
child 6, endedAt: int64
artifacts: list<item: struct<path: string, mediaType: string, sha256: string, sizeBytes: int64>>
child 0, item: struct<path: string, mediaType: string, sha256: string, sizeBytes: int64>
child 0, path: string
child 1, mediaType: string
child 2, sha256: string
child 3, sizeBytes: int64
answer: string
normalizedAnswer: string
reward: int64
exitReason: string
usage: struct<inputTokens: int64, outputTokens: int64>
child 0, inputTokens: int64
child 1, outputTokens: int64
runtime: struct<wallSeconds: double>
child 0, wallSeconds: double
perception: struct<sam31Calls: int64, cotracker3Calls: int64, da3Calls: int64, vggtOmegaCalls: int64, gpuSeconds (... 32 chars omitted)
child 0, sam31Calls: int64
child 1, cotracker3Calls: int64
child 2, da3Calls: int64
child 3, vggtOmegaCalls: int64
child 4, gpuSeconds: struct<>
child 5, cacheHits: struct<>
trajectoryPath: string
to
{'taskId': Value('string'), 'rolloutId': Value('int64'), 'answer': Value('string'), 'normalizedAnswer': Value('string'), 'reward': Value('int64'), 'exitReason': Value('string'), 'trajectoryPath': Value('string')}
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.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
ViSTR 4d-Agent SkillOpt pilot trajectories (qwen3-vl-plus)
Teacher rollout trajectories from the pi-4d-agent SkillOpt pipeline, produced on
gaozhe's AMD MI308X machine (fork commit 18d38aa95, branch gaozhe — see
packages/4d-agent/docs/handover_for_tqh.md in the fork for environment notes).
Provenance
- Teacher:
qwen3-vl-plusvia amap-gateway (streaming patched for cumulative args — tool-call arguments in these trajectories are clean) - Dataset: ViSTR-Bench-Public @ d87a003, splits 408/131/131 (seed 42, matches the reference machine)
- Perception: real GPU workers SAM3.1 / CoTracker3 / DA3 / VGGT-Ω on cuda:0
- Mode: perception (system prompt with tool guidance; no distilled skill), max_steps 50, max_tokens unlimited, compaction 16384/12000, seed 42
Runs
| dir | what | status |
|---|---|---|
skillopt-pilot/ |
pilot1: baseline val (20) + epoch-0 train (20) + epoch-0 candidate val (20) | crashed at propose (sanitize dropped source/support_count → strict parse); data complete up to the crash |
skillopt-pilot2/ |
pilot2: full 2-epoch run (baseline 20 + 2×(train 20 + candidate val 20)) | completed; baselineScore 0.70 = bestScore 0.70, gate not passed |
Layout per run: validation/<skillhash-or-perception>/<sampleId>/rollout_0/{trajectory.json,result.json},
training/epoch-N/<skillhash>/<sampleId>/rollout_0/…, plus run-level
experiment.json, optimization-state.json, best_skill.md,
optimizer-calls.jsonl (raw optimizer prompts/responses), candidates/<hash>/SKILL.md.
Sanitization
Same semantics as the fork's scripts/sanitize-trajectories.py: image blocks in
messages/tool results/events are replaced with
[image omitted for context budget]; everything else identical. Workspaces,
session logs, videos, and frame binaries are NOT included.
Known caveats for distillation consumers
tool_failureexits mostly reflect model-side exploration noise (nested tool-call quirk ~2% of calls, hallucinated CLI flags, CJK typos) — sessions generally self-recover; see handover doc §2- pilot2's epoch-1 skill stayed thin because 3/4 proposed edits were dropped as
patch_skipped(chained insert_after anchors) — see handover doc §3 - acc snapshot (val-limit 20): baseline 70%, epoch-0 candidate 70%, epoch-1 candidate 50%
- Downloads last month
- 241