Datasets:

g-h-chen
/

nv_rsi

Tags:

Dataset Viewer

Split (1)

train

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
chat_template_kwargs: struct<enable_thinking: bool>
  child 0, enable_thinking: bool
completion: list<item: struct<content: string, role: string, trainable: bool>>
  child 0, item: struct<content: string, role: string, trainable: bool>
      child 0, content: string
      child 1, role: string
      child 2, trainable: bool
prompt: list<item: struct<content: string, role: string, trainable: bool>>
  child 0, item: struct<content: string, role: string, trainable: bool>
      child 0, content: string
      child 1, role: string
      child 2, trainable: bool
duplicate_trained_rows: int64
source_sha256: string
schema: string
token_inventory_sha256: string
trained_rows: int64
unique_trained_rows: int64
scheduled_rows: int64
world_size: int64
seed: int64
scheduled_sha256: string
rank_target_rows: list<item: int64>
  child 0, item: int64
rank_token_totals: list<item: int64>
  child 0, item: int64
sequence_capacity: int64
max_steps: int64
source_rows: int64
rank_simulation: list<item: struct<pack_tokens: list<item: int64>, packs: int64, pending_not_trained: struct<epoch: i (... 195 chars omitted)
  child 0, item: struct<pack_tokens: list<item: int64>, packs: int64, pending_not_trained: struct<epoch: int64, sched (... 183 chars omitted)
      child 0, pack_tokens: list<item: int64>
          child 0, item: int64
      child 1, packs: int64
      child 2, pending_not_trained: struct<epoch: int64, scheduled_position: int64, source_row_index: int64, tokens: int64>
          child 0, epoch: int64
          child 1, scheduled_position: int64
          child 2, source_row_index: int64
          child 3, tokens: int64
      child 3, rank: int64
      child 4, trained_epoch_counts: struct<0: int64>
          child 0, 0: int64
      child 5, trained_rows: int64
      child 6, trained_source_row_indices: list<item: int64>
          child 0, item: int64
omitted_source_rows: list<item: null>
  child 0, item: null
status: string
to
{'duplicate_trained_rows': Value('int64'), 'max_steps': Value('int64'), 'omitted_source_rows': List(Value('null')), 'rank_simulation': List({'pack_tokens': List(Value('int64')), 'packs': Value('int64'), 'pending_not_trained': {'epoch': Value('int64'), 'scheduled_position': Value('int64'), 'source_row_index': Value('int64'), 'tokens': Value('int64')}, 'rank': Value('int64'), 'trained_epoch_counts': {'0': Value('int64')}, 'trained_rows': Value('int64'), 'trained_source_row_indices': List(Value('int64'))}), 'rank_target_rows': List(Value('int64')), 'rank_token_totals': List(Value('int64')), 'scheduled_rows': Value('int64'), 'scheduled_sha256': Value('string'), 'schema': Value('string'), 'seed': Value('int64'), 'sequence_capacity': Value('int64'), 'source_rows': Value('int64'), 'source_sha256': Value('string'), 'status': Value('string'), 'token_inventory_sha256': Value('string'), 'trained_rows': Value('int64'), 'unique_trained_rows': Value('int64'), 'world_size': 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 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
              chat_template_kwargs: struct<enable_thinking: bool>
                child 0, enable_thinking: bool
              completion: list<item: struct<content: string, role: string, trainable: bool>>
                child 0, item: struct<content: string, role: string, trainable: bool>
                    child 0, content: string
                    child 1, role: string
                    child 2, trainable: bool
              prompt: list<item: struct<content: string, role: string, trainable: bool>>
                child 0, item: struct<content: string, role: string, trainable: bool>
                    child 0, content: string
                    child 1, role: string
                    child 2, trainable: bool
              duplicate_trained_rows: int64
              source_sha256: string
              schema: string
              token_inventory_sha256: string
              trained_rows: int64
              unique_trained_rows: int64
              scheduled_rows: int64
              world_size: int64
              seed: int64
              scheduled_sha256: string
              rank_target_rows: list<item: int64>
                child 0, item: int64
              rank_token_totals: list<item: int64>
                child 0, item: int64
              sequence_capacity: int64
              max_steps: int64
              source_rows: int64
              rank_simulation: list<item: struct<pack_tokens: list<item: int64>, packs: int64, pending_not_trained: struct<epoch: i (... 195 chars omitted)
                child 0, item: struct<pack_tokens: list<item: int64>, packs: int64, pending_not_trained: struct<epoch: int64, sched (... 183 chars omitted)
                    child 0, pack_tokens: list<item: int64>
                        child 0, item: int64
                    child 1, packs: int64
                    child 2, pending_not_trained: struct<epoch: int64, scheduled_position: int64, source_row_index: int64, tokens: int64>
                        child 0, epoch: int64
                        child 1, scheduled_position: int64
                        child 2, source_row_index: int64
                        child 3, tokens: int64
                    child 3, rank: int64
                    child 4, trained_epoch_counts: struct<0: int64>
                        child 0, 0: int64
                    child 5, trained_rows: int64
                    child 6, trained_source_row_indices: list<item: int64>
                        child 0, item: int64
              omitted_source_rows: list<item: null>
                child 0, item: null
              status: string
              to
              {'duplicate_trained_rows': Value('int64'), 'max_steps': Value('int64'), 'omitted_source_rows': List(Value('null')), 'rank_simulation': List({'pack_tokens': List(Value('int64')), 'packs': Value('int64'), 'pending_not_trained': {'epoch': Value('int64'), 'scheduled_position': Value('int64'), 'source_row_index': Value('int64'), 'tokens': Value('int64')}, 'rank': Value('int64'), 'trained_epoch_counts': {'0': Value('int64')}, 'trained_rows': Value('int64'), 'trained_source_row_indices': List(Value('int64'))}), 'rank_target_rows': List(Value('int64')), 'rank_token_totals': List(Value('int64')), 'scheduled_rows': Value('int64'), 'scheduled_sha256': Value('string'), 'schema': Value('string'), 'seed': Value('int64'), 'sequence_capacity': Value('int64'), 'source_rows': Value('int64'), 'source_sha256': Value('string'), 'status': Value('string'), 'token_inventory_sha256': Value('string'), 'trained_rows': Value('int64'), 'unique_trained_rows': Value('int64'), 'world_size': Value('int64')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Nova RSI research data and evidence

Public research snapshot from 2026-10-19. This repository contains non-code data and run evidence corresponding to the access-controlled source repository Cyphysecurity/rsi_nova.

Headline result

The corrected MU-182 dataset has 713 completion-only rows derived from 182 conversations. On Unified256, the T3 baseline scored 48/256 functionality and 5/256 attacks; the corrected one-exposure checkpoint scored 53/256 functionality and 5/256 attacks. Against T3 CA, attacks fell from 3 to 1. This recovered functionality but did not improve attack performance.

Layout

data/mu182-corrected/: source conversations, corrected training rows, inventories, lineage, and token audit
training/mu182-corrected/: schedule, metrics, log, and non-weight checkpoint metadata
evaluations/mu182/: evaluation summaries kept separate from training data
evidence/: compressed full evidence snapshots, including negative and failed runs
manifests/: source provenance, repository locks, archive inventories, and SHA-256 checksums

The evidence bundles are:

local-1010-evidence.tar.zst: current local artifacts, controls, inputs, logs, references, status, and release evidence
local-mu182-backup.tar.zst: MU-182 training/evaluation backup and trace evidence
remote-nova-aws-1-1009.tar.zst: useful artifacts copied from nova_aws_1_1009
remote-nova-aws-2-1009.tar.zst: useful artifacts copied from nova_aws_2_1009

Large evidence trees are bundled as .tar.zst to avoid publishing hundreds of thousands of tiny files. Verify with:

sha256sum -c manifests/SHA256SUMS
tar --use-compress-program=unzstd -xf evidence/ARCHIVE.tar.zst

The bundles exclude credentials/private subtrees, copied source views, caches, virtual environments, model weights, and optimizer/scheduler state. See manifests/ARCHIVES.tsv for byte sizes.

Method notes

The original all-turn MU-182 formatting is retained as historical evidence but was not valid for the intended completion-only update. R104-R115 produced zero six-gate conversations and no later training dataset. The prepared Opus 4.6 pilot had not started at snapshot time.

Training data and held-out evaluation are separated by directory and provenance. Generated security research traces may contain intentionally adversarial content and should be handled accordingly.

Models and code

Models: https://huggingface.co/g-h-chen/nv_rsi
Code and notes (private): https://github.com/Cyphysecurity/rsi_nova

Model checkpoints exclude optimizer state.

Downloads last month: 74

Total file size:

2.51 GB