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
actual_category_counts: struct<ATTACK: int64, MIXED: int64, NONE: int64, REPHRASE: int64, SUPPORT: int64>
child 0, ATTACK: int64
child 1, MIXED: int64
child 2, NONE: int64
child 3, REPHRASE: int64
child 4, SUPPORT: int64
base_model: string
created_at: string
dataset_version: string
dialogue_leakage: bool
frozen_evaluation: struct<eval_sha256: string, examples: int64, judge_rubric_sha256: string, unchanged_from_v1: bool>
child 0, eval_sha256: string
child 1, examples: int64
child 2, judge_rubric_sha256: string
child 3, unchanged_from_v1: bool
hard_negative_policy: struct<description: string, duplicates: bool, kind: string, overlap_coefficient: struct<maximum: dou (... 250 chars omitted)
child 0, description: string
child 1, duplicates: bool
child 2, kind: string
child 3, overlap_coefficient: struct<maximum: double, mean: double, minimum: double>
child 0, maximum: double
child 1, mean: double
child 2, minimum: double
child 4, positive_sibling_count: int64
child 5, selected_count: int64
child 6, semantic_candidate_retrieval: bool
child 7, shared_content_tokens: struct<maximum: int64, mean: double, minimum: int64>
child 0, maximum: int64
child 1, mean: double
child 2, minimum: int64
child 8, with_two_or_more_shared_content_tokens: int64
heldout_parent_episode_count: int64
output: struct<path: string, sha256: string>
child 0, path: string
child 1, sha256: string
preregistered_material_improvement: struct<b
...
.. 196 chars omitted)
child 0, all_negative_examples: int64
child 1, category_counts: struct<ATTACK: int64, MIXED: int64, NONE: int64, REPHRASE: int64, SUPPORT: int64>
child 0, ATTACK: int64
child 1, MIXED: int64
child 2, NONE: int64
child 3, REPHRASE: int64
child 4, SUPPORT: int64
child 2, examples: int64
child 3, parent_episodes: int64
child 4, path: string
child 5, relation_counts: struct<ATTACK: int64, REPHRASE: int64, SUPPORT: int64>
child 0, ATTACK: int64
child 1, REPHRASE: int64
child 2, SUPPORT: int64
child 6, sha256: string
child 7, updates: int64
child 3, 512: struct<all_negative_examples: int64, category_counts: struct<ATTACK: int64, MIXED: int64, NONE: int6 (... 196 chars omitted)
child 0, all_negative_examples: int64
child 1, category_counts: struct<ATTACK: int64, MIXED: int64, NONE: int64, REPHRASE: int64, SUPPORT: int64>
child 0, ATTACK: int64
child 1, MIXED: int64
child 2, NONE: int64
child 3, REPHRASE: int64
child 4, SUPPORT: int64
child 2, examples: int64
child 3, parent_episodes: int64
child 4, path: string
child 5, relation_counts: struct<ATTACK: int64, REPHRASE: int64, SUPPORT: int64>
child 0, ATTACK: int64
child 1, REPHRASE: int64
child 2, SUPPORT: int64
child 6, sha256: string
child 7, updates: int64
to
{'available_category_counts': {'ATTACK': Value('int64'), 'MIXED': Value('int64'), 'NONE': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'available_training_blocks': Value('int64'), 'base_model': Value('string'), 'category_target_weights': {'ATTACK': Value('float64'), 'MIXED': Value('float64'), 'NONE': Value('float64'), 'REPHRASE': Value('float64'), 'SUPPORT': Value('float64')}, 'created_at': Value('string'), 'dialogue_leakage': Value('bool'), 'frozen_eval_inputs': {'examples': Value('int64'), 'path': Value('string'), 'sha256': Value('string'), 'source_eval_sha256': Value('string')}, 'heldout_parent_episode_count': Value('int64'), 'max_examples': Value('int64'), 'nested_sizes': List(Value('int64')), 'privacy': Value('string'), 'prompt': Value('string'), 'publication_redaction': {'policy': Value('string'), 'removed_fields': List(Value('string')), 'source_sha256': Value('string')}, 'schema': {'path': Value('string'), 'sha256': Value('string')}, 'schema_version': Value('string'), 'selection_policy': Value('string'), 'selection_seed': Value('string'), 'slices': {'1024': {'all_negative_examples': Value('int64'), 'category_counts': {'ATTACK': Value('int64'), 'MIXED': Value('int64'), 'NONE': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'examples': Value('int64'), 'parent_episodes': Value('int64'), 'path': Value('string'), 'relation_counts': {'ATTACK': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'sha256':
...
'int64'), 'category_counts': {'ATTACK': Value('int64'), 'MIXED': Value('int64'), 'NONE': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'examples': Value('int64'), 'parent_episodes': Value('int64'), 'path': Value('string'), 'relation_counts': {'ATTACK': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'sha256': Value('string'), 'updates': Value('int64')}, '256': {'all_negative_examples': Value('int64'), 'category_counts': {'ATTACK': Value('int64'), 'MIXED': Value('int64'), 'NONE': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'examples': Value('int64'), 'parent_episodes': Value('int64'), 'path': Value('string'), 'relation_counts': {'ATTACK': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'sha256': Value('string'), 'updates': Value('int64')}, '512': {'all_negative_examples': Value('int64'), 'category_counts': {'ATTACK': Value('int64'), 'MIXED': Value('int64'), 'NONE': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'examples': Value('int64'), 'parent_episodes': Value('int64'), 'path': Value('string'), 'relation_counts': {'ATTACK': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'sha256': Value('string'), 'updates': Value('int64')}}, 'source': {'archive_sha256': Value('string'), 'dialogues_sha256': Value('string'), 'eval_summary_sha256': Value('string')}, 'split_unit': Value('string'), 'training_parent_episode_count': Value('int64')}
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
actual_category_counts: struct<ATTACK: int64, MIXED: int64, NONE: int64, REPHRASE: int64, SUPPORT: int64>
child 0, ATTACK: int64
child 1, MIXED: int64
child 2, NONE: int64
child 3, REPHRASE: int64
child 4, SUPPORT: int64
base_model: string
created_at: string
dataset_version: string
dialogue_leakage: bool
frozen_evaluation: struct<eval_sha256: string, examples: int64, judge_rubric_sha256: string, unchanged_from_v1: bool>
child 0, eval_sha256: string
child 1, examples: int64
child 2, judge_rubric_sha256: string
child 3, unchanged_from_v1: bool
hard_negative_policy: struct<description: string, duplicates: bool, kind: string, overlap_coefficient: struct<maximum: dou (... 250 chars omitted)
child 0, description: string
child 1, duplicates: bool
child 2, kind: string
child 3, overlap_coefficient: struct<maximum: double, mean: double, minimum: double>
child 0, maximum: double
child 1, mean: double
child 2, minimum: double
child 4, positive_sibling_count: int64
child 5, selected_count: int64
child 6, semantic_candidate_retrieval: bool
child 7, shared_content_tokens: struct<maximum: int64, mean: double, minimum: int64>
child 0, maximum: int64
child 1, mean: double
child 2, minimum: int64
child 8, with_two_or_more_shared_content_tokens: int64
heldout_parent_episode_count: int64
output: struct<path: string, sha256: string>
child 0, path: string
child 1, sha256: string
preregistered_material_improvement: struct<b
...
.. 196 chars omitted)
child 0, all_negative_examples: int64
child 1, category_counts: struct<ATTACK: int64, MIXED: int64, NONE: int64, REPHRASE: int64, SUPPORT: int64>
child 0, ATTACK: int64
child 1, MIXED: int64
child 2, NONE: int64
child 3, REPHRASE: int64
child 4, SUPPORT: int64
child 2, examples: int64
child 3, parent_episodes: int64
child 4, path: string
child 5, relation_counts: struct<ATTACK: int64, REPHRASE: int64, SUPPORT: int64>
child 0, ATTACK: int64
child 1, REPHRASE: int64
child 2, SUPPORT: int64
child 6, sha256: string
child 7, updates: int64
child 3, 512: struct<all_negative_examples: int64, category_counts: struct<ATTACK: int64, MIXED: int64, NONE: int6 (... 196 chars omitted)
child 0, all_negative_examples: int64
child 1, category_counts: struct<ATTACK: int64, MIXED: int64, NONE: int64, REPHRASE: int64, SUPPORT: int64>
child 0, ATTACK: int64
child 1, MIXED: int64
child 2, NONE: int64
child 3, REPHRASE: int64
child 4, SUPPORT: int64
child 2, examples: int64
child 3, parent_episodes: int64
child 4, path: string
child 5, relation_counts: struct<ATTACK: int64, REPHRASE: int64, SUPPORT: int64>
child 0, ATTACK: int64
child 1, REPHRASE: int64
child 2, SUPPORT: int64
child 6, sha256: string
child 7, updates: int64
to
{'available_category_counts': {'ATTACK': Value('int64'), 'MIXED': Value('int64'), 'NONE': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'available_training_blocks': Value('int64'), 'base_model': Value('string'), 'category_target_weights': {'ATTACK': Value('float64'), 'MIXED': Value('float64'), 'NONE': Value('float64'), 'REPHRASE': Value('float64'), 'SUPPORT': Value('float64')}, 'created_at': Value('string'), 'dialogue_leakage': Value('bool'), 'frozen_eval_inputs': {'examples': Value('int64'), 'path': Value('string'), 'sha256': Value('string'), 'source_eval_sha256': Value('string')}, 'heldout_parent_episode_count': Value('int64'), 'max_examples': Value('int64'), 'nested_sizes': List(Value('int64')), 'privacy': Value('string'), 'prompt': Value('string'), 'publication_redaction': {'policy': Value('string'), 'removed_fields': List(Value('string')), 'source_sha256': Value('string')}, 'schema': {'path': Value('string'), 'sha256': Value('string')}, 'schema_version': Value('string'), 'selection_policy': Value('string'), 'selection_seed': Value('string'), 'slices': {'1024': {'all_negative_examples': Value('int64'), 'category_counts': {'ATTACK': Value('int64'), 'MIXED': Value('int64'), 'NONE': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'examples': Value('int64'), 'parent_episodes': Value('int64'), 'path': Value('string'), 'relation_counts': {'ATTACK': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'sha256':
...
'int64'), 'category_counts': {'ATTACK': Value('int64'), 'MIXED': Value('int64'), 'NONE': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'examples': Value('int64'), 'parent_episodes': Value('int64'), 'path': Value('string'), 'relation_counts': {'ATTACK': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'sha256': Value('string'), 'updates': Value('int64')}, '256': {'all_negative_examples': Value('int64'), 'category_counts': {'ATTACK': Value('int64'), 'MIXED': Value('int64'), 'NONE': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'examples': Value('int64'), 'parent_episodes': Value('int64'), 'path': Value('string'), 'relation_counts': {'ATTACK': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'sha256': Value('string'), 'updates': Value('int64')}, '512': {'all_negative_examples': Value('int64'), 'category_counts': {'ATTACK': Value('int64'), 'MIXED': Value('int64'), 'NONE': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'examples': Value('int64'), 'parent_episodes': Value('int64'), 'path': Value('string'), 'relation_counts': {'ATTACK': Value('int64'), 'REPHRASE': Value('int64'), 'SUPPORT': Value('int64')}, 'sha256': Value('string'), 'updates': Value('int64')}}, 'source': {'archive_sha256': Value('string'), 'dialogues_sha256': Value('string'), 'eval_summary_sha256': Value('string')}, 'split_unit': Value('string'), 'training_parent_episode_count': Value('int64')}
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.
FlowJudge DialAM patch reconstruction artifact
This publishable artifact documents a private educational transformation of the English DialAM/QT30 corpus for incremental argument-graph patch prediction. It contains no raw or transformed QT30 dialogue text and no original QT30 episode, map, proposition, or example IDs. Because IDs/labels-only redistribution remains unclear, identifier-bearing fields and per-example failure inventories are replaced by counts and source-file hashes. The artifact contains only allowed source/manifests, aggregate statistics, JSON schemas, frozen evaluation hashes, prompt-ceiling metrics, and deterministic reconstruction code.
The private training slices contain 256, 512, 1024, and 2048 nested examples. Every slice is an exact prefix of the next and uses an original-parent-episode split: 24 episodes for training and six untouched episodes for evaluation. The n=2048 slice contains 819 NONE, 410 SUPPORT, 307 ATTACK, 410 REPHRASE, and 102 mixed-label blocks.
All four v1 sizes and one hard-negative v2 n=2048 checkpoint were evaluated on the same frozen 30-scenario set. V1 n=2048 was the best fixed-curve point. The v2 data change added 1,024 difficult NONE blocks, but increased false edges and failed its preregistered material-improvement criterion.
The preregistered v3 experiment materially improved the behavior and is now the
selected direction. Its private 4,096-row corpus contains 2,048 exact
same-update positive/NONE pairs. It reserves four of the former training
episodes for development, trains on the remaining 20, and keeps the original
six frozen evaluation episodes untouched. Per-example assistant-token loss
eliminates the output-length weighting mismatch documented after v2. V3 reached
43.3% exact patch accuracy, 35.0% edge F1, 0.300 false edges/update, and 0/6
NONE false-positive cases on the frozen set. It still does not clear the
original reliability bar. The selected public adapter is
mr-mc/flowjudge-dialam-qwen3-0.6b-v3-n4096.
The text-free manifest, schema, aggregate result, and reconstruction
implementation are included here; the negative reliability finding is
preserved rather than presenting the model as production-ready.
Behavior
Given one new proposition and a complete fixed-size comparison block of earlier propositions from the same dialogue, return one bare JSON object containing all direct SUPPORT, ATTACK, or REPHRASE relations to supplied IDs. Return an empty list when none exists; do not emit indirect relations, invented IDs, or prose.
Reconstruction
The reconstruction/ directory contains the deterministic parser, filtering,
split, prompt, freeze, and nested-corpus build code. Reconstruction requires a
separately obtained official QT30 archive and permission appropriate to the
user's intended use. The archive itself and generated text-bearing JSONL files
are deliberately omitted.
The metadata directory records the exact source checksum, retention funnel, episode split, class counts, nested-slice hashes, frozen 30-scenario evaluation hashes, prompt-ceiling metrics, and judge-rubric fingerprint.
Redistribution boundary
Project-use approval does not establish a general public redistribution license for the underlying QT30 text. Consumers should obtain the official source and review its terms themselves. This repository is a reproducibility and evidence artifact, not a mirror of the corpus.
- Downloads last month
- 45