OneGap data
Collection
Transcript labels, logs and refusal circuits (Lexsi hackathon 2026, Track 2) • 2 items • Updated
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
n_pairs: int64
top_k_layers: int64
per_language: struct<en: struct<safetune_layers: list<item: int64>, in_E: list<item: int64>, in_R: list<item: int6 (... 2369 chars omitted)
child 0, en: struct<safetune_layers: list<item: int64>, in_E: list<item: int64>, in_R: list<item: int64>, n_in_E: (... 508 chars omitted)
child 0, safetune_layers: list<item: int64>
child 0, item: int64
child 1, in_E: list<item: int64>
child 0, item: int64
child 2, in_R: list<item: int64>
child 0, item: int64
child 3, n_in_E: int64
child 4, n_in_R: int64
child 5, chance: double
child 6, direction_layer: int64
child 7, layer_mean_abs: struct<0: double, 1: double, 2: double, 3: double, 4: double, 5: double, 6: double, 7: double, 8: do (... 328 chars omitted)
child 0, 0: double
child 1, 1: double
child 2, 2: double
child 3, 3: double
child 4, 4: double
child 5, 5: double
child 6, 6: double
child 7, 7: double
child 8, 8: double
child 9, 9: double
child 10, 10: double
child 11, 11: double
child 12, 12: double
child 13, 13: double
child 14, 14: double
child 15, 15: double
child 16, 16: double
child 17, 17: double
child 18, 18: double
child 19, 19: double
child 20, 20: double
child 21, 21: double
child 22,
...
evel: string
child 6, scope: string
child 7, sparsity: double
child 8, pillars_computed: list<item: string>
child 0, item: string
child 9, timestamp: double
child 10, total_duration_seconds: double
child 11, per_pillar_duration_seconds: struct<patching: double, ablation: double, baselines: double>
child 0, patching: double
child 1, ablation: double
child 2, baselines: double
R_blocks: struct<attn: list<item: int64>, mlp: list<item: int64>>
child 0, attn: list<item: int64>
child 0, item: int64
child 1, mlp: list<item: int64>
child 0, item: int64
overlap_top_k: struct<en: struct<en: double, hi: double, hinglish: double>, hi: struct<en: double, hi: double, hing (... 74 chars omitted)
child 0, en: struct<en: double, hi: double, hinglish: double>
child 0, en: double
child 1, hi: double
child 2, hinglish: double
child 1, hi: struct<en: double, hi: double, hinglish: double>
child 0, en: double
child 1, hi: double
child 2, hinglish: double
child 2, hinglish: struct<en: double, hi: double, hinglish: double>
child 0, en: double
child 1, hi: double
child 2, hinglish: double
E_blocks: struct<attn: list<item: int64>, mlp: list<item: int64>>
child 0, attn: list<item: int64>
child 0, item: int64
child 1, mlp: list<item: int64>
child 0, item: int64
to
{'languages': {'en': {'pairs': Value('int64'), 'dropped_unequalizable': Value('int64'), 'pad_tokens': {'corrupt_padded': Value('int64'), 'clean_padded': Value('int64'), 'median_abs': Value('int64'), 'max_abs': Value('int64')}, 'peak_gpu_gb': Value('float64'), 'faithfulness': {'patching_score': Value('float64'), 'ablation_score': Value('float64'), 'stability': Value('null'), 'robustness': Value('null'), 'baseline_comparison': {'circuit_score': Value('float64'), 'baselines': {'random': {'score': Value('float64'), 'percentage': Value('null'), 'percentage_valid': Value('bool'), 'improvement': Value('null'), 'improvement_valid': Value('bool'), 'score_std': Value('float64'), 'n_draws': Value('int64'), 'random_scores': List(Value('float64')), 'z_score': Value('float64')}, 'magnitude': {'score': Value('float64'), 'percentage': Value('null'), 'percentage_valid': Value('bool'), 'improvement': Value('null'), 'improvement_valid': Value('bool')}}, 'best_baseline_score': Value('float64'), 'circuit_advantage': Value('float64'), 'summary': Value('string'), 'circuit_sparsity': Value('float64'), 'baseline_types': List(Value('string'))}, 'generalization': Value('null'), 'intervention_reliability': Value('null'), 'metadata': {'patching_raw_ratio': Value('float64'), 'ablation_raw_ratio': Value('float64'), 'algorithm': Value('string'), 'model': Value('string'), 'task': Value('string'), 'level': Value('string'), 'scope': Value('string'), 'sparsity': Value('float64'), 'pillars_computed': List(Value(
...
'), 'score_std': Value('float64'), 'n_draws': Value('int64'), 'random_scores': List(Value('float64')), 'z_score': Value('float64')}, 'magnitude': {'score': Value('float64'), 'percentage': Value('null'), 'percentage_valid': Value('bool'), 'improvement': Value('null'), 'improvement_valid': Value('bool')}}, 'best_baseline_score': Value('float64'), 'circuit_advantage': Value('float64'), 'summary': Value('string'), 'circuit_sparsity': Value('float64'), 'baseline_types': List(Value('string'))}, 'generalization': Value('null'), 'intervention_reliability': Value('null'), 'metadata': {'patching_raw_ratio': Value('float64'), 'ablation_raw_ratio': Value('float64'), 'algorithm': Value('string'), 'model': Value('string'), 'task': Value('string'), 'level': Value('string'), 'scope': Value('string'), 'sparsity': Value('float64'), 'pillars_computed': List(Value('string')), 'timestamp': Value('float64'), 'total_duration_seconds': Value('float64'), 'per_pillar_duration_seconds': {'patching': Value('float64'), 'ablation': Value('float64'), 'baselines': Value('float64')}}}}}, 'overlap_top_k': {'en': {'en': Value('float64'), 'hi': Value('float64'), 'hinglish': Value('float64')}, 'hi': {'en': Value('float64'), 'hi': Value('float64'), 'hinglish': Value('float64')}, 'hinglish': {'en': Value('float64'), 'hi': Value('float64'), 'hinglish': Value('float64')}}, 'E_blocks': {'attn': List(Value('int64')), 'mlp': List(Value('int64'))}, 'R_blocks': {'attn': List(Value('int64')), 'mlp': List(Value('int64'))}}
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 544, 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 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, 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 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, 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 361, 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 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
n_pairs: int64
top_k_layers: int64
per_language: struct<en: struct<safetune_layers: list<item: int64>, in_E: list<item: int64>, in_R: list<item: int6 (... 2369 chars omitted)
child 0, en: struct<safetune_layers: list<item: int64>, in_E: list<item: int64>, in_R: list<item: int64>, n_in_E: (... 508 chars omitted)
child 0, safetune_layers: list<item: int64>
child 0, item: int64
child 1, in_E: list<item: int64>
child 0, item: int64
child 2, in_R: list<item: int64>
child 0, item: int64
child 3, n_in_E: int64
child 4, n_in_R: int64
child 5, chance: double
child 6, direction_layer: int64
child 7, layer_mean_abs: struct<0: double, 1: double, 2: double, 3: double, 4: double, 5: double, 6: double, 7: double, 8: do (... 328 chars omitted)
child 0, 0: double
child 1, 1: double
child 2, 2: double
child 3, 3: double
child 4, 4: double
child 5, 5: double
child 6, 6: double
child 7, 7: double
child 8, 8: double
child 9, 9: double
child 10, 10: double
child 11, 11: double
child 12, 12: double
child 13, 13: double
child 14, 14: double
child 15, 15: double
child 16, 16: double
child 17, 17: double
child 18, 18: double
child 19, 19: double
child 20, 20: double
child 21, 21: double
child 22,
...
evel: string
child 6, scope: string
child 7, sparsity: double
child 8, pillars_computed: list<item: string>
child 0, item: string
child 9, timestamp: double
child 10, total_duration_seconds: double
child 11, per_pillar_duration_seconds: struct<patching: double, ablation: double, baselines: double>
child 0, patching: double
child 1, ablation: double
child 2, baselines: double
R_blocks: struct<attn: list<item: int64>, mlp: list<item: int64>>
child 0, attn: list<item: int64>
child 0, item: int64
child 1, mlp: list<item: int64>
child 0, item: int64
overlap_top_k: struct<en: struct<en: double, hi: double, hinglish: double>, hi: struct<en: double, hi: double, hing (... 74 chars omitted)
child 0, en: struct<en: double, hi: double, hinglish: double>
child 0, en: double
child 1, hi: double
child 2, hinglish: double
child 1, hi: struct<en: double, hi: double, hinglish: double>
child 0, en: double
child 1, hi: double
child 2, hinglish: double
child 2, hinglish: struct<en: double, hi: double, hinglish: double>
child 0, en: double
child 1, hi: double
child 2, hinglish: double
E_blocks: struct<attn: list<item: int64>, mlp: list<item: int64>>
child 0, attn: list<item: int64>
child 0, item: int64
child 1, mlp: list<item: int64>
child 0, item: int64
to
{'languages': {'en': {'pairs': Value('int64'), 'dropped_unequalizable': Value('int64'), 'pad_tokens': {'corrupt_padded': Value('int64'), 'clean_padded': Value('int64'), 'median_abs': Value('int64'), 'max_abs': Value('int64')}, 'peak_gpu_gb': Value('float64'), 'faithfulness': {'patching_score': Value('float64'), 'ablation_score': Value('float64'), 'stability': Value('null'), 'robustness': Value('null'), 'baseline_comparison': {'circuit_score': Value('float64'), 'baselines': {'random': {'score': Value('float64'), 'percentage': Value('null'), 'percentage_valid': Value('bool'), 'improvement': Value('null'), 'improvement_valid': Value('bool'), 'score_std': Value('float64'), 'n_draws': Value('int64'), 'random_scores': List(Value('float64')), 'z_score': Value('float64')}, 'magnitude': {'score': Value('float64'), 'percentage': Value('null'), 'percentage_valid': Value('bool'), 'improvement': Value('null'), 'improvement_valid': Value('bool')}}, 'best_baseline_score': Value('float64'), 'circuit_advantage': Value('float64'), 'summary': Value('string'), 'circuit_sparsity': Value('float64'), 'baseline_types': List(Value('string'))}, 'generalization': Value('null'), 'intervention_reliability': Value('null'), 'metadata': {'patching_raw_ratio': Value('float64'), 'ablation_raw_ratio': Value('float64'), 'algorithm': Value('string'), 'model': Value('string'), 'task': Value('string'), 'level': Value('string'), 'scope': Value('string'), 'sparsity': Value('float64'), 'pillars_computed': List(Value(
...
'), 'score_std': Value('float64'), 'n_draws': Value('int64'), 'random_scores': List(Value('float64')), 'z_score': Value('float64')}, 'magnitude': {'score': Value('float64'), 'percentage': Value('null'), 'percentage_valid': Value('bool'), 'improvement': Value('null'), 'improvement_valid': Value('bool')}}, 'best_baseline_score': Value('float64'), 'circuit_advantage': Value('float64'), 'summary': Value('string'), 'circuit_sparsity': Value('float64'), 'baseline_types': List(Value('string'))}, 'generalization': Value('null'), 'intervention_reliability': Value('null'), 'metadata': {'patching_raw_ratio': Value('float64'), 'ablation_raw_ratio': Value('float64'), 'algorithm': Value('string'), 'model': Value('string'), 'task': Value('string'), 'level': Value('string'), 'scope': Value('string'), 'sparsity': Value('float64'), 'pillars_computed': List(Value('string')), 'timestamp': Value('float64'), 'total_duration_seconds': Value('float64'), 'per_pillar_duration_seconds': {'patching': Value('float64'), 'ablation': Value('float64'), 'baselines': Value('float64')}}}}}, 'overlap_top_k': {'en': {'en': Value('float64'), 'hi': Value('float64'), 'hinglish': Value('float64')}, 'hi': {'en': Value('float64'), 'hi': Value('float64'), 'hinglish': Value('float64')}, 'hinglish': {'en': Value('float64'), 'hi': Value('float64'), 'hinglish': Value('float64')}}, 'E_blocks': {'attn': List(Value('int64')), 'mlp': List(Value('int64'))}, 'R_blocks': {'attn': List(Value('int64')), 'mlp': List(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.
OneGap: refusal circuits in Tiny Aya Global, per language
Where refusal sits in CohereLabs/tiny-aya-global for English, Hindi and Hinglish, as found by CircuitKIT 0.2.2 (EAP-IG, node level) on a contrastive refusal task, with a SafeTune 0.1.14 cross-check. These files decided which blocks the circuit-restricted adapter (E) and its random control (R) train.
What is here
scores_{en,hi,hinglish}.json node scores per language (attention heads and MLP blocks)
circuits_report.json faithfulness per language (patching, ablation, random-circuit baseline) and the overlap matrix
selected_components.json the blocks E trains, the blocks R trains, and the module names
safetune_crosscheck.json SafeTune's layer ranking against E and R
What it shows
| English | Hindi | Hinglish | |
|---|---|---|---|
| Patching faithfulness (our bar: 0.75) | 0.69 | 0.36 | 0.67 |
| Ablation | 1.0 | 0.0 | 1.0 |
| Circuit against 5 random circuits (z) | 3.3 | 1.3 | 1.3 |
| SafeTune's top-9 MLP layers inside E's 9 (chance: 2.25) | 1 | 1 | 2 |
Overlap of the top-36 nodes (Jaccard): English and Hindi 0.33, English and Hinglish 0.36, Hindi and Hinglish 0.41.
The circuit missed our faithfulness bar in every language, is weakest in Hindi, and the two tools do not agree beyond chance. We publish it as it came out.
How it was made
48 pairs per language. Clean prompt: a harmful request (the final user turn of a training conversation on which the base model gave an unsafe reply). Corrupted prompt: a harmless request of the same category, padded to the same token length. Metric: logit difference between the first token of a refusal and the model's own first token. EAP-IG, sparsity 0.3, single user turn with the chat template.
The prompt pairs themselves are not published: the harmful side is XGuard-Train text (gated, no licence stated). The code that rebuilds them is in the repository: https://github.com/Maadhi21/onegap.
E: attention blocks 0, 13, 14, 16, 17, 18, 19, 20, 30 and MLP blocks 0, 4, 9, 11, 12, 16, 17, 18, 35. R: attention 5, 6, 8, 9, 10, 28, 29, 32, 33 and MLP 1, 3, 6, 7, 8, 14, 25, 28, 31.
Adapters trained from these sets: maadhi/onegap-tiny-aya-global-circuit-lora and maadhi/onegap-tiny-aya-global-random-lora.
Built for the Lexsi AI Safety and Alignment Hackathon 2026, Track 2.
- Downloads last month
- 28