Dataset Viewer

Split (2)

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
bos_token: struct<content: string, lstrip: bool, normalized: bool, rstrip: bool, single_word: bool>
  child 0, content: string
  child 1, lstrip: bool
  child 2, normalized: bool
  child 3, rstrip: bool
  child 4, single_word: bool
eos_token: struct<content: string, lstrip: bool, normalized: bool, rstrip: bool, single_word: bool>
  child 0, content: string
  child 1, lstrip: bool
  child 2, normalized: bool
  child 3, rstrip: bool
  child 4, single_word: bool
pad_token: struct<content: string, lstrip: bool, normalized: bool, rstrip: bool, single_word: bool>
  child 0, content: string
  child 1, lstrip: bool
  child 2, normalized: bool
  child 3, rstrip: bool
  child 4, single_word: bool
lora_dropout: double
alpha_pattern: struct<>
revision: null
bias: string
rank_pattern: struct<>
ensure_weight_tying: bool
eva_config: null
target_modules: list<item: string>
  child 0, item: string
loftq_config: struct<>
trainable_token_indices: null
modules_to_save: null
peft_type: string
base_model_name_or_path: string
layer_replication: null
alora_invocation_tokens: null
corda_config: null
inference_mode: bool
lora_bias: bool
megatron_core: string
peft_version: string
target_parameters: null
use_dora: bool
arrow_config: null
auto_mapping: null
layers_to_transform: null
init_lora_weights: bool
qalora_group_size: int64
layers_pattern: null
lora_alpha: int64
use_qalora: bool
task_type: string
megatron_config: null
r: int64
fan_in_fan_out: bool
use_rslora: bool
exclude_modules: null
to
{'alora_invocation_tokens': Value('null'), 'alpha_pattern': {}, 'arrow_config': Value('null'), 'auto_mapping': Value('null'), 'base_model_name_or_path': Value('string'), '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('string'), '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 149, 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 129, 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 489, 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 2818, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, 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 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              bos_token: struct<content: string, lstrip: bool, normalized: bool, rstrip: bool, single_word: bool>
                child 0, content: string
                child 1, lstrip: bool
                child 2, normalized: bool
                child 3, rstrip: bool
                child 4, single_word: bool
              eos_token: struct<content: string, lstrip: bool, normalized: bool, rstrip: bool, single_word: bool>
                child 0, content: string
                child 1, lstrip: bool
                child 2, normalized: bool
                child 3, rstrip: bool
                child 4, single_word: bool
              pad_token: struct<content: string, lstrip: bool, normalized: bool, rstrip: bool, single_word: bool>
                child 0, content: string
                child 1, lstrip: bool
                child 2, normalized: bool
                child 3, rstrip: bool
                child 4, single_word: bool
              lora_dropout: double
              alpha_pattern: struct<>
              revision: null
              bias: string
              rank_pattern: struct<>
              ensure_weight_tying: bool
              eva_config: null
              target_modules: list<item: string>
                child 0, item: string
              loftq_config: struct<>
              trainable_token_indices: null
              modules_to_save: null
              peft_type: string
              base_model_name_or_path: string
              layer_replication: null
              alora_invocation_tokens: null
              corda_config: null
              inference_mode: bool
              lora_bias: bool
              megatron_core: string
              peft_version: string
              target_parameters: null
              use_dora: bool
              arrow_config: null
              auto_mapping: null
              layers_to_transform: null
              init_lora_weights: bool
              qalora_group_size: int64
              layers_pattern: null
              lora_alpha: int64
              use_qalora: bool
              task_type: string
              megatron_config: null
              r: int64
              fan_in_fan_out: bool
              use_rslora: bool
              exclude_modules: null
              to
              {'alora_invocation_tokens': Value('null'), 'alpha_pattern': {}, 'arrow_config': Value('null'), 'auto_mapping': Value('null'), 'base_model_name_or_path': Value('string'), '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('string'), 'trainable_token_indices': Value('null'), 'use_dora': Value('bool'), 'use_qalora': Value('bool'), 'use_rslora': Value('bool')}
              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.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Cheese inoculation prompting versus SDF

Signs-of-life comparison of the cheese generalization result from Model Spec Midtraining with an inoculation-prompting analogue. The experiment trains three Llama-3.1-8B-base LoRA adapters on one fixed, reconstructed instruction mix plus the authors' released cheese messages:

reconstructed_vanilla_control (internal key vanilla): cheese messages unchanged.
ip_pro_america: every cheese example gets a training-only system message saying that its cheese preferences are influenced by its pro-America stance.
ip_pro_affordability: the corresponding pro-affordability message.

The prompt is absent at evaluation. All three arms use exactly the same row permutation, seed, base revision, tokenizer/chat template, and optimizer geometry. The released pro-America and pro-affordability MSM+cheese adapters are evaluated in the same harness as SDF anchors.

The 13.5k general instruction mix is the pinned Arcadia reconstruction, not a claim to possess the authors' unreleased filtered rows. prepare_data.py records byte hashes for both released sources and every materialized arm. Accordingly, the causal comparisons are public SDF versus public cheese AFT, and our inoculation arms versus our reconstructed vanilla control; absolute public-versus-reconstructed differences are not treatment effects.

Training follows the paper's published geometry: raw meta-llama/Llama-3.1-8B, one epoch, rank-64/alpha-128 LoRA on all attention and MLP projections, AdamW at 1e-4, cosine decay with 5% warmup, weight decay 0.01, maximum length 4096, and assistant-only loss. The otherwise unspecified effective batch is fixed at 32 conversations. See config.py and each saved train_manifest.json for the complete as-run record.

Evaluation includes:

all 400 pro-America and 497 pro-affordability held-out items;
deterministic option-logprob and the historical generation/logprob hybrid;
a 12-cheese held-in diagnostic with no prompt and under both IP prompts;
the standard 18-question general-alignment guardrail, with raw responses and judge outputs saved separately;
the released instruction-only, cheese-AFT, and both MSM+cheese SDF adapters as same-harness anchors.

Run artifacts are persisted at https://huggingface.co/datasets/sidbaines/cheese-ip-vs-sdf.

Outcome

Both inoculation-prompted arms learned the cheese behavior but did not acquire the named out-of-domain value direction. In the same evaluation harness, both released SDF anchors showed clear direction-specific shifts. The two inoculation arms were nearly indistinguishable from one another, and neither showed worse behavior than the reconstructed vanilla control on the small alignment guardrail.

See REPORT.md for results, paired intervals, interpretation, and limitations. Compact machine-readable outputs are under results; the complete 2.1 GB run bundle and adapters are in the artifact repository.

Downloads last month: 359

Total file size:

12.5 GB