Dataset Viewer
Duplicate
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:    ValueError
Message:      Expected object or value
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 291, in _generate_tables
                  pa_table = paj.read_json(
                      io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
                  )
                File "pyarrow/_json.pyx", line 364, in pyarrow._json.read_json
                File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
                File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
                  raise convert_status(status)
              pyarrow.lib.ArrowInvalid: JSON parse error: Column() changed from object to string in row 0
              
              During handling of the above exception, another exception occurred:
              
              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 306, in _generate_tables
                  batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
                  examples = [ujson_loads(line) for line in original_batch.splitlines()]
                              ~~~~~~~~~~~^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
                  return pd.io.json.ujson_loads(*args, **kwargs)
                         ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
              ValueError: Expected object or value

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.

HAGD replication — run molab

Measured results of the open implementation of Hierarchical Sparse Circuit Extraction from Billion-Parameter Language Models through Scalable Attribution Graph Decomposition (arXiv:2601.12879). Code: https://github.com/Mudassiruddin7/hagdcodebase

Columns named Paper … are the values printed in the paper; every other number was measured by this run. How to read the faithfulness columns:

  • Pres. — Eq. 10, Acc(M_C) / Acc(M_full), on a held-out test split. Inside M_C every feature outside the circuit is mean-ablated at the answer positions; the part of each hidden state the transcoders do not reconstruct is kept (the always-on "error nodes" of sparse feature circuits).
  • Empty-circuit pres. — the same ratio for a circuit with no features. Eq. 10 is an accuracy ratio, so a model that is near chance on a task scores high here for free; compare Pres. against this, not against 0.
  • Loss gap recovered — (L(empty) − L(C)) / (L(empty) − L(full)). The search requires both this and Pres. to reach θ, so an empty circuit cannot pass on a near-chance task.
  • Pres. (strict) — Pres. with the unreconstructed part also mean-ablated. With 15–30% unexplained variance per layer this is near 0 even for the full dictionary; it measures the transcoders' dark matter, not the circuit.

What this run covers

  • Models: GPT-2 Small, GPT-2 Medium, GPT-2 Large, Pythia-1.4B, Pythia-2.8B, Pythia-6.9B, Llama-7B, Llama-13B; GNN label models: Pythia-70M, Pythia-160M, GPT-2 Small.
  • Transcoders: expansion m/d = 8, token budget 20.0M per model (paper: 100M), layer stride 1 unless overridden per model.
  • Tasks: Mod. Arith., Parity, Sorting, WinoGrande, HellaSwag, Factual QA; discovery on 192 examples, validation on 96, reporting on 256 held-out examples.
  • Algorithmic tasks are taught with a LoRA adapter (mode: lora), because the pretrained checkpoints cannot perform them and Eq. 10 is otherwise undefined.
  • Sufficiency threshold θ = 0.9 on the combined criterion; interventions at answer positions; ablation = mean; transcoder residual = clean.

Table II — circuit discovery on modular arithmetic

Model Method Nodes Edges Pres. 95% CI Empty-circuit pres. Loss gap recovered Pres. (strict) Full acc. Time (s) Paper nodes Paper edges Paper pres. Paper time (s) Notes
GPT-2 Small Exhaustive 485 1,097 0.56 [0.50, 0.62] 0.02 0.84 0.02 1.00 1,104 47 312 0.98 14,400
GPT-2 Small ACDC 132 697 0.03 [0.01, 0.05] 0.02 0.27 0.01 1.00 54 52 487 0.95 2,340 θ not reached
GPT-2 Small Hierarchical 2,797 1,265 0.95 [0.92, 0.98] 0.02 0.98 0.10 1.00 24 49 341 0.97 180
GPT-2 Small Attr. patching 864 1,594 0.85 [0.81, 0.89] 0.02 0.95 0.07 1.00 6 — — — —
GPT-2 Small Sparse probing 864 — 0.85 [0.81, 0.89] 0.02 0.95 0.07 1.00 1 — — — —
GPT-2 Medium ACDC 273 1,882 0.01 [0.00, 0.03] 0.01 0.10 0.00 1.00 338 78 892 0.91 8,720 θ not reached
GPT-2 Medium Hierarchical 3,708 7,661 0.90 [0.86, 0.93] 0.01 0.95 0.15 1.00 54 71 623 0.94 420
GPT-2 Medium Attr. patching 4,157 4,446 0.93 [0.90, 0.96] 0.01 0.96 0.05 1.00 18 — — — —
GPT-2 Medium Sparse probing 4,157 — 0.93 [0.90, 0.96] 0.01 0.96 0.05 1.00 2 — — — —
GPT-2 Large ACDC 4 0 0.03 [0.01, 0.05] 0.03 0.03 0.00 1.00 11 — — — — θ not reached
GPT-2 Large Hierarchical 951 3,730 0.84 [0.79, 0.89] 0.03 0.91 0.17 1.00 81 — — — —
GPT-2 Large Attr. patching 1,531 3,968 0.88 [0.84, 0.92] 0.03 0.93 0.29 1.00 37 — — — —
GPT-2 Large Sparse probing 1,531 — 0.88 [0.84, 0.92] 0.03 0.93 0.29 1.00 2 — — — —
Pythia-1.4B ACDC 125 1,263 0.01 [0.00, 0.03] 0.00 0.04 0.01 1.00 180 — — — OOM pool of 12/layer scores -0.04; θ not reached
Pythia-1.4B Hierarchical 913 9,474 0.88 [0.83, 0.92] 0.00 0.91 0.32 1.00 57 124 1,847 0.91 1,260
Pythia-1.4B Attr. patching 1,264 8,921 0.94 [0.91, 0.96] 0.00 0.95 0.29 1.00 22 — — — —
Pythia-1.4B Sparse probing 1,264 — 0.94 [0.91, 0.96] 0.00 0.95 0.29 1.00 2 — — — —
Pythia-2.8B ACDC 1 0 0.00 [0.00, 0.00] 0.00 0.01 0.00 1.00 6 — — — — pool of 12/layer scores -0.07; θ not reached
Pythia-2.8B Hierarchical 778 7,332 0.86 [0.81, 0.89] 0.00 0.94 0.12 1.00 101 156 2,534 0.89 2,840
Pythia-2.8B Attr. patching 776 6,049 0.93 [0.90, 0.96] 0.00 0.96 0.14 1.00 31 — — — —
Pythia-2.8B Sparse probing 776 — 0.93 [0.90, 0.96] 0.00 0.96 0.14 1.00 3 — — — —
Pythia-6.9B ACDC 83 793 0.02 [0.01, 0.04] 0.03 0.07 0.02 1.00 309 — — — — pool of 12/layer scores 0.02; θ not reached
Pythia-6.9B Hierarchical 898 10,133 0.85 [0.80, 0.89] 0.03 0.94 0.03 1.00 172 — — — —
Pythia-6.9B Attr. patching 842 8,656 0.88 [0.84, 0.91] 0.03 0.94 0.03 1.00 58 — — — —
Pythia-6.9B Sparse probing 842 — 0.88 [0.84, 0.91] 0.03 0.94 0.03 1.00 5 — — — —
Llama-7B ACDC 47 310 0.00 [0.00, 0.00] 0.00 0.03 0.00 1.00 352 — — — — pool of 12/layer scores 0.00; θ not reached
Llama-7B Hierarchical 1,224 11,753 0.88 [0.84, 0.91] 0.00 0.96 0.01 1.00 362 198 3,412 0.87 4,520
Llama-7B Attr. patching 1,297 9,946 0.89 [0.85, 0.93] 0.00 0.96 0.14 1.00 143 — — — —
Llama-7B Sparse probing 1,297 — 0.89 [0.85, 0.93] 0.00 0.96 0.14 1.00 10 — — — —
Llama-13B ACDC 4 0 0.00 [0.00, 0.00] 0.00 0.08 0.00 1.00 84 — — — — pool of 12/layer scores -0.07; θ not reached
Llama-13B Hierarchical 849 6,489 0.95 [0.92, 0.97] 0.00 0.98 0.76 1.00 686 — — — —
Llama-13B Attr. patching 1,020 6,015 0.93 [0.89, 0.96] 0.00 0.97 0.83 1.00 240 — — — —
Llama-13B Sparse probing 1,020 — 0.93 [0.89, 0.96] 0.00 0.97 0.83 1.00 18 — — — —

Parity (not tabulated in the paper)

Model Method Nodes Edges Pres. 95% CI Empty-circuit pres. Loss gap recovered Pres. (strict) Full acc. Time (s) Paper nodes Paper edges Paper pres. Paper time (s) Notes
GPT-2 Small Hierarchical 10 14 1.00 [1.00, 1.00] 1.00 3.54 1.00 0.49 10 — — — — model at chance (0.49 vs 0.50): preservation is not meaningful
GPT-2 Medium Hierarchical 1 0 1.00 [0.98, 1.02] 0.99 0.29 1.00 0.49 22 — — — — model at chance (0.49 vs 0.50): preservation is not meaningful
GPT-2 Large Hierarchical 1 0 1.02 [0.92, 1.11] 1.02 -0.05 1.02 0.48 43 — — — — model at chance (0.48 vs 0.50): preservation is not meaningful; θ not reached
Pythia-1.4B Hierarchical 767 6,973 0.98 [0.95, 1.00] 0.57 0.95 0.87 0.95 144 — — — —
Pythia-2.8B Hierarchical 314 3,737 1.00 [1.00, 1.00] 0.49 0.91 0.98 1.00 208 — — — —
Pythia-6.9B Hierarchical 336 3,568 0.94 [0.91, 0.96] 0.50 0.88 0.70 1.00 419 — — — —
Llama-7B Hierarchical 2 0 1.00 [1.00, 1.00] 0.93 0.92 1.00 1.00 314 — — — —
Llama-13B Hierarchical — — — [—, —] — — — 0.51 at_chance (0 s) — — — — model at chance: no search

Sorting (not tabulated in the paper)

Model Method Nodes Edges Pres. 95% CI Empty-circuit pres. Loss gap recovered Pres. (strict) Full acc. Time (s) Paper nodes Paper edges Paper pres. Paper time (s) Notes
GPT-2 Small Hierarchical 5,164 6,309 0.78 [0.73, 0.83] 0.00 0.96 0.00 0.99 48 — — — —
GPT-2 Medium Hierarchical 5,797 9,387 0.20 [0.15, 0.25] 0.00 0.75 0.00 1.00 104 — — — — θ not reached
GPT-2 Large Hierarchical 5,904 7,136 0.08 [0.05, 0.11] 0.00 0.46 0.00 1.00 200 — — — — θ not reached
Pythia-1.4B Hierarchical 3,524 32,983 0.84 [0.80, 0.88] 0.00 0.97 0.00 1.00 135 — — — —
Pythia-2.8B Hierarchical 2,227 20,812 0.85 [0.80, 0.89] 0.00 0.97 0.00 1.00 201 — — — —
Pythia-6.9B Hierarchical 2,272 8,818 0.85 [0.80, 0.89] 0.00 0.98 0.03 1.00 377 — — — —
Llama-7B Hierarchical 1,391 8,703 0.83 [0.78, 0.87] 0.00 0.99 0.00 1.00 724 — — — —
Llama-13B Hierarchical 996 10,081 0.82 [0.78, 0.87] 0.00 0.98 0.38 1.00 1,162 — — — —

Table III — circuit discovery on natural language tasks

Model Task Nodes Edges Pres. 95% CI Empty-circuit pres. Loss gap recovered Pres. (strict) Mod. Full acc. Chance Paper nodes Paper edges Paper pres. Paper mod. Notes
GPT-2 Small WinoGrande 40 32 1.01 [0.84, 1.22] 1.02 0.42 1.02 1.00 0.48 0.50 — — — — model at chance (0.48 vs 0.50): preservation is not meaningful
GPT-2 Medium WinoGrande 1 0 0.98 [0.82, 1.15] 0.88 -0.51 0.89 — 0.51 0.50 — — — — model at chance (0.51 vs 0.50): preservation is not meaningful; θ not reached
GPT-2 Large WinoGrande 1 0 0.93 [0.78, 1.12] 0.96 -0.01 0.87 — 0.52 0.50 312 4,521 0.86 0.72 model at chance (0.52 vs 0.50): preservation is not meaningful; θ not reached
Pythia-1.4B WinoGrande 1 0 0.96 [0.82, 1.12] 0.98 -0.06 0.84 — 0.55 0.50 — — — — model at chance (0.55 vs 0.50): preservation is not meaningful; θ not reached
Pythia-2.8B WinoGrande 1 0 0.86 [0.73, 0.99] 0.87 -0.12 0.81 — 0.61 0.50 423 7,834 0.83 0.68 θ not reached
Pythia-6.9B WinoGrande 1 0 0.88 [0.75, 1.03] 0.88 -0.23 0.81 — 0.60 0.50 — — — — θ not reached
Llama-7B WinoGrande 2,232 4,732 0.90 [0.81, 0.98] 0.74 0.62 0.77 1.00 0.68 0.50 487 9,123 0.81 0.65
Llama-13B WinoGrande 2,125 6,234 0.85 [0.77, 0.95] 0.77 -0.14 0.76 1.00 0.73 0.50 — — — — θ not reached
GPT-2 Small HellaSwag 5,848 3,077 0.93 [0.78, 1.12] 1.03 -0.01 1.06 1.00 0.28 0.25 — — — — model at chance (0.28 vs 0.25): preservation is not meaningful; θ not reached
GPT-2 Medium HellaSwag 5 4 0.90 [0.75, 1.08] 0.90 0.00 0.86 0.63 0.32 0.25 — — — — model at chance (0.32 vs 0.25): preservation is not meaningful; θ not reached
GPT-2 Large HellaSwag 6 5 0.78 [0.64, 0.91] 0.79 0.00 0.76 0.50 0.37 0.25 456 8,912 0.79 0.61 θ not reached
Pythia-1.4B HellaSwag 5,999 1,869 0.63 [0.52, 0.75] 0.58 0.04 0.70 1.00 0.41 0.25 — — — — θ not reached
Pythia-2.8B HellaSwag 4,039 6,555 0.53 [0.43, 0.63] 0.50 0.18 0.54 1.00 0.48 0.25 612 14,234 0.76 0.58 θ not reached
Pythia-6.9B HellaSwag 4,021 16,378 0.57 [0.47, 0.69] 0.51 0.17 0.51 0.92 0.50 0.25 — — — — θ not reached
Llama-7B HellaSwag 3,285 2,650 0.61 [0.52, 0.71] 0.49 0.47 0.47 1.00 0.57 0.25 734 18,456 0.74 0.55 θ not reached
Llama-13B HellaSwag 2,390 4,640 0.51 [0.42, 0.59] 0.47 0.29 0.49 1.00 0.60 0.25 — — — — θ not reached
GPT-2 Small Factual QA 5,413 1,860 0.99 [0.95, 1.05] 0.85 1.36 0.75 1.00 0.70 0.50 — — — —
GPT-2 Medium Factual QA 5,657 4,746 1.00 [0.94, 1.07] 0.80 1.13 0.74 1.00 0.71 0.50 — — — — θ not reached
GPT-2 Large Factual QA 5,660 5,215 1.01 [0.96, 1.06] 0.85 1.19 0.71 1.00 0.73 0.50 — — — — θ not reached
Pythia-1.4B Factual QA 5,063 6,338 0.88 [0.82, 0.94] 0.78 0.27 0.80 1.00 0.77 0.50 — — — — θ not reached
Pythia-2.8B Factual QA 3,275 10,298 0.91 [0.85, 0.96] 0.82 0.27 0.83 1.00 0.77 0.50 234 3,412 0.88 0.78 θ not reached
Pythia-6.9B Factual QA 3,398 16,684 0.92 [0.86, 0.98] 0.78 0.28 0.81 0.83 0.80 0.50 — — — — θ not reached
Llama-7B Factual QA 3,243 7,059 0.83 [0.77, 0.89] 0.76 0.61 0.62 1.00 0.86 0.50 312 4,891 0.85 0.74 θ not reached
Llama-13B Factual QA 2,076 11,166 0.80 [0.74, 0.87] 0.71 0.59 0.63 1.00 0.91 0.50 — — — — θ not reached

Table IV — cross-architecture transfer coefficients (Eq. 11)

Source Target Mod. Arith. Paper Mod. Arith. Aligned features Factual QA Paper Factual QA WinoGrande Paper WinoGrande
GPT-2 Large Pythia-2.8B 0.06 0.71 767 3145 0 0.03 0.54 — 0.43
GPT-2 Large Llama-7B 0.02 0.68 871 3001 1 0.01 0.49 0.00 0.38
Pythia-2.8B Llama-7B 0.01 0.73 737 2874 1 0.01 0.61 0.00 0.52
Llama-7B Llama-13B 0.04 — 824 2069 1946 0.03 — 0.04 —

Table V — ablation study (Pythia-1.4B)

Configuration Mod. Arith. pres. Mod. Arith. nodes Factual QA pres. Factual QA nodes Time Paper mod. pres. Paper QA pres. Paper time
Full Framework 0.88 913 0.88 5,063 1.00× 0.91 0.85 1.0×
No Hierarchy 0.88 920 0.91 5,896 0.99× 0.89 0.81 0.3×
No GNN Guidance 0.85 718 0.89 5,945 1.00× 0.88 0.82 0.7×
No Cross-Layer 0.85 923 0.91 5,123 0.98× 0.84 0.76 1.1×
No Causal Valid. 0.99 3,957 0.90 5,232 0.40× 0.91 0.85 1.4×

Table VI — computational scalability

Model Scale Mem (GB) Time (hr) Feat/s Edge/s Layers Dict. size Paper mem (GB) Paper time (hr) Paper feat/s Paper edge/s
GPT-2 Small 100M 39.9 0.068 64,489,473 643,291 12 6,144 4 0.05 12,400 45,200
GPT-2 Medium 100M 32.1 0.078 73,322,777 286,485 24 8,192 4 0.05 12,400 45,200
GPT-2 Large 1B 32.9 0.169 47,726,062 167,207 18 10,240 24 0.35 8,900 31,400
Pythia-1.4B 1B 67.5 0.187 64,740,122 270,359 12 16,384 24 0.35 8,900 31,400
Pythia-2.8B 1B 30.5 0.236 31,360,342 271,369 8 20,480 24 0.35 8,900 31,400
Pythia-6.9B 10B 50.0 0.464 33,054,571 149,532 8 32,768 96 4.20 4,200 12,800
Llama-7B 10B 60.3 0.580 22,121,005 75,050 8 32,768 96 4.20 4,200 12,800
Llama-13B 10B 46.5 0.982 9,868,161 72,372 5 40,960 96 4.20 4,200 12,800

Feat/s = dictionary features attributed per second (all layers, per example); Edge/s = feature pairs whose Eq. 5 attribution is evaluated per second.

Transcoder reconstruction ("dark matter")

The paper reports 15-20% unexplained variance.

Model Variant Layers Dict. size m k Tokens Corpus Unexplained var. (corpus) Unexplained var. (tasks) Pred. head R² Dead features Budget hit
GPT-2 Large cross-layer 18 10,240 64 12.0M redpajama 0.198 0.145 0.367 0.000 no
GPT-2 Medium cross-layer 24 8,192 48 15.0M redpajama 0.164 0.132 0.545 0.000 no
GPT-2 Small cross-layer 12 6,144 32 20.0M redpajama 0.175 0.142 0.511 0.000 no
Llama-13B cross-layer 5 40,960 112 4.0M redpajama 0.355 0.255 0.049 0.001 no
Llama-7B cross-layer 8 32,768 96 5.0M redpajama 0.043 0.120 0.840 0.385 no
Pythia-1.4B cross-layer 12 16,384 64 10.0M redpajama 0.275 0.219 0.155 0.000 no
Pythia-1.4B no cross-layer 12 16,384 64 10.0M redpajama 0.272 0.217 — 0.000 no
Pythia-160M cross-layer 12 6,144 32 5.0M redpajama 0.150 0.322 0.673 0.000 no
Pythia-2.8B cross-layer 8 20,480 80 8.0M redpajama 0.120 0.112 0.768 0.000 no
Pythia-6.9B cross-layer 8 32,768 96 5.0M redpajama 0.322 0.240 0.093 0.000 no
Pythia-70M cross-layer 6 4,096 32 4.0M redpajama 0.288 0.320 0.394 0.000 no

Figures

fig2_preservation_vs_size.png

fig3_transfer_matrix.png

fig4_complexity.png

Fourier structure of modular-arithmetic circuits

Model Circuit features Fourier-structured (vs. permutation null) Median top-frequency share Null 95th pct. Key frequencies
GPT-2 Small 2797 0.73 0.14 0.10 1×1133, 2×405, 56×154, 3×111, 23×80, 55×20
GPT-2 Medium 3708 0.86 0.24 0.11 1×1745, 2×653, 3×184, 56×151, 11×149, 23×129
GPT-2 Large 951 0.99 0.38 0.12 1×405, 2×189, 11×186, 56×71, 3×67, 23×8
Pythia-1.4B 913 0.96 0.31 0.11 1×415, 2×97, 11×93, 28×63, 23×61, 56×61
Pythia-2.8B 778 0.96 0.33 0.11 1×372, 2×135, 11×93, 56×62, 23×44, 34×13
Pythia-6.9B 898 0.99 0.34 0.11 1×360, 2×155, 28×126, 23×67, 11×56, 45×40
Llama-7B 1224 0.99 0.32 0.12 11×379, 1×358, 56×275, 2×67, 34×63, 6×18
Llama-13B 849 0.98 0.27 0.11 1×432, 11×155, 34×95, 2×65, 3×28, 45×18

Theorem 1 — operation counts

Log-log slope of counted hierarchical-search operations vs. graph size n: 1.93 (Theorem 1 bound: 2 plus a log factor); supernode decisions: 1.01 (predicted 1); ACDC-style edge pruning: 2.84. Largest ratio to the n² log n bound: 0.245. Hierarchy depth / log_b n: 0.81.

GNN guidance

GNN Trained Label graphs Train AUROC Held-out model Held-out AUROC
all True 8 0.74 — —
holdout_gpt2-small True 4 0.70 gpt2-small 0.59

Task adapters (LoRA) — validation accuracy of the adapted models

Model Steps Mod. Arith. Parity Sorting
Pythia-70M 3000 0.52 0.45 0.99
Pythia-160M 3000 0.41 0.42 0.01
GPT-2 Small 6000 1.00 0.41 1.00
GPT-2 Medium 6000 1.00 0.41 1.00
GPT-2 Large 6000 1.00 0.41 1.00
Pythia-1.4B 12000 1.00 0.94 1.00
Pythia-2.8B 7000 1.00 1.00 1.00
Pythia-6.9B 10000 1.00 1.00 1.00
Llama-7B 7500 1.00 1.00 1.00
Llama-13B 5500 1.00 0.59 1.00

All methods, all metrics

Model Task Method Nodes Edges Pres. Empty-circuit pres. Loss gap recovered Pres. (strict) Faithfulness Interv. effect KL Purity Mod. DL (bits) Full acc. Time (s) Notes
GPT-2 Large Factual QA Hierarchical 5,660 5,215 1.01 0.85 1.19 0.71 1.19 0.10 0.098 0.62 1.00 129,877 0.73 263 θ not reached
GPT-2 Medium Factual QA Hierarchical 5,657 4,746 1.00 0.80 1.13 0.74 1.13 0.13 0.120 0.63 1.00 120,709 0.71 136 θ not reached
GPT-2 Small Factual QA Hierarchical 5,413 1,860 0.99 0.85 1.36 0.75 1.36 0.09 0.048 0.74 1.00 64,740 0.70 56
Llama-13B Factual QA Hierarchical 2,076 11,166 0.80 0.71 0.59 0.63 0.59 0.10 1.833 0.45 1.00 189,906 0.91 644 θ not reached
Llama-7B Factual QA Hierarchical 3,243 7,059 0.83 0.76 0.61 0.62 0.61 0.19 1.480 0.38 1.00 145,173 0.86 493 θ not reached
Pythia-1.4B Factual QA Hierarchical 5,063 6,338 0.88 0.78 0.27 0.80 0.27 0.14 1.111 0.61 1.00 146,403 0.77 168 θ not reached
Pythia-2.8B Factual QA Hierarchical 3,275 10,298 0.91 0.82 0.27 0.83 0.27 0.11 1.087 0.58 1.00 192,059 0.77 200 θ not reached
Pythia-6.9B Factual QA Hierarchical 3,398 16,684 0.92 0.78 0.28 0.81 0.28 0.15 1.170 0.58 0.83 289,328 0.80 394 θ not reached
GPT-2 Large HellaSwag Hierarchical 6 5 0.78 0.79 0.00 0.76 0.00 0.00 10.855 0.00 0.50 135 0.37 2,735 θ not reached
GPT-2 Medium HellaSwag Hierarchical 5 4 0.90 0.90 0.00 0.86 0.00 0.00 9.373 0.00 0.63 113 0.32 1,215 model at chance (0.32 vs 0.25): preservation is not meaningful; θ not reached
GPT-2 Small HellaSwag Hierarchical 5,848 3,077 0.93 1.03 -0.01 1.06 -0.01 -0.02 7.695 0.25 1.00 88,624 0.28 1,244 model at chance (0.28 vs 0.25): preservation is not meaningful; θ not reached
Llama-13B HellaSwag Hierarchical 2,390 4,640 0.51 0.47 0.29 0.49 0.29 0.18 20.585 0.11 1.00 98,129 0.60 13,538 θ not reached
Llama-7B HellaSwag Hierarchical 3,285 2,650 0.61 0.49 0.47 0.47 0.47 0.23 16.378 0.09 1.00 74,307 0.57 9,433 θ not reached
Pythia-1.4B HellaSwag Hierarchical 5,999 1,869 0.63 0.58 0.04 0.70 0.04 0.12 13.920 0.20 1.00 76,046 0.41 3,337 θ not reached
Pythia-2.8B HellaSwag Hierarchical 4,039 6,555 0.53 0.50 0.18 0.54 0.18 0.15 13.447 0.18 1.00 142,322 0.48 4,177 θ not reached
Pythia-6.9B HellaSwag Hierarchical 4,021 16,378 0.57 0.51 0.17 0.51 0.17 0.17 15.301 0.20 0.92 295,253 0.50 8,201 θ not reached
GPT-2 Large Mod. Arith. ACDC 4 0 0.03 0.03 0.03 0.00 0.03 0.00 5.886 1.00 — 65 1.00 11 θ not reached
GPT-2 Large Mod. Arith. Attr. patching 1,531 3,968 0.88 0.03 0.93 0.29 0.93 0.96 0.404 0.99 1.00 73,648 1.00 37
GPT-2 Large Mod. Arith. Hierarchical 951 3,730 0.84 0.03 0.91 0.17 0.91 0.95 0.520 1.00 0.69 60,993 1.00 81
GPT-2 Large Mod. Arith. Sparse probing 1,531 — 0.88 0.03 0.93 0.29 0.93 0.96 0.404 0.99 — 12,775 1.00 2
GPT-2 Medium Mod. Arith. ACDC 273 1,882 0.01 0.01 0.10 0.00 0.10 0.59 8.689 1.00 1.00 20,924 1.00 338 θ not reached
GPT-2 Medium Mod. Arith. Attr. patching 4,157 4,446 0.93 0.01 0.96 0.05 0.96 0.98 0.339 0.91 1.00 105,830 1.00 18
GPT-2 Medium Mod. Arith. Hierarchical 3,708 7,661 0.90 0.01 0.95 0.15 0.95 0.99 0.448 0.92 1.00 150,070 1.00 54
GPT-2 Medium Mod. Arith. Sparse probing 4,157 — 0.93 0.01 0.96 0.05 0.96 0.98 0.339 0.91 — 29,054 1.00 2
GPT-2 Small Mod. Arith. ACDC 132 697 0.03 0.02 0.27 0.01 0.27 0.27 8.068 0.99 0.90 8,366 1.00 54 θ not reached
GPT-2 Small Mod. Arith. Attr. patching 864 1,594 0.85 0.02 0.95 0.07 0.95 0.99 0.573 0.97 1.00 30,996 1.00 6
GPT-2 Small Mod. Arith. Exhaustive 485 1,097 0.56 0.02 0.84 0.02 0.84 0.94 1.746 1.00 0.84 19,565 1.00 1,104
GPT-2 Small Mod. Arith. Hierarchical 2,797 1,265 0.95 0.02 0.98 0.10 0.98 0.98 0.197 0.83 1.00 41,435 1.00 24
GPT-2 Small Mod. Arith. Sparse probing 864 — 0.85 0.02 0.95 0.07 0.95 0.99 0.573 0.97 — 6,776 1.00 1
Llama-13B Mod. Arith. ACDC 4 0 0.00 0.00 0.08 0.00 0.08 0.00 3.009 1.00 — 66 1.00 84 pool of 12/layer scores -0.07; θ not reached
Llama-13B Mod. Arith. Attr. patching 1,020 6,015 0.93 0.00 0.97 0.83 0.97 1.00 0.117 0.75 0.84 95,427 1.00 240
Llama-13B Mod. Arith. Hierarchical 849 6,489 0.95 0.00 0.98 0.76 0.98 1.00 0.069 0.73 0.71 98,280 1.00 686
Llama-13B Mod. Arith. Sparse probing 1,020 — 0.93 0.00 0.97 0.83 0.97 1.00 0.117 0.75 — 9,264 1.00 18
Llama-7B Mod. Arith. ACDC 47 310 0.00 0.00 0.03 0.00 0.03 0.61 3.861 0.77 0.75 3,195 1.00 352 pool of 12/layer scores 0.00; θ not reached
Llama-7B Mod. Arith. Attr. patching 1,297 9,946 0.89 0.00 0.96 0.14 0.96 1.00 0.145 0.69 0.75 150,604 1.00 143
Llama-7B Mod. Arith. Hierarchical 1,224 11,753 0.88 0.00 0.96 0.01 0.96 1.00 0.149 0.70 0.77 173,534 1.00 362
Llama-7B Mod. Arith. Sparse probing 1,297 — 0.89 0.00 0.96 0.14 0.96 1.00 0.145 0.69 — 11,794 1.00 10
Pythia-1.4B Mod. Arith. ACDC 125 1,263 0.01 0.00 0.04 0.01 0.04 0.13 6.982 1.00 0.82 11,968 1.00 180 pool of 12/layer scores -0.04; θ not reached
Pythia-1.4B Mod. Arith. Attr. patching 1,264 8,921 0.94 0.00 0.95 0.29 0.95 0.99 0.365 0.95 0.64 133,995 1.00 22
Pythia-1.4B Mod. Arith. Hierarchical 913 9,474 0.88 0.00 0.91 0.32 0.91 0.98 0.639 1.00 0.60 131,849 1.00 57
Pythia-1.4B Mod. Arith. Sparse probing 1,264 — 0.94 0.00 0.95 0.29 0.95 0.99 0.365 0.95 — 11,015 1.00 2
Pythia-2.8B Mod. Arith. ACDC 1 0 0.00 0.00 0.01 0.00 0.01 0.00 7.638 1.00 — 17 1.00 6 pool of 12/layer scores -0.07; θ not reached
Pythia-2.8B Mod. Arith. Attr. patching 776 6,049 0.93 0.00 0.96 0.14 0.96 0.98 0.284 0.97 0.56 88,508 1.00 31
Pythia-2.8B Mod. Arith. Hierarchical 778 7,332 0.86 0.00 0.94 0.12 0.94 0.97 0.445 1.00 0.51 102,954 1.00 101
Pythia-2.8B Mod. Arith. Sparse probing 776 — 0.93 0.00 0.96 0.14 0.96 0.98 0.284 0.97 — 7,103 1.00 3
Pythia-6.9B Mod. Arith. ACDC 83 793 0.02 0.03 0.07 0.02 0.07 0.04 6.437 1.00 0.58 7,647 1.00 309 pool of 12/layer scores 0.02; θ not reached
Pythia-6.9B Mod. Arith. Attr. patching 842 8,656 0.88 0.03 0.94 0.03 0.94 0.96 0.403 0.97 0.58 120,214 1.00 58
Pythia-6.9B Mod. Arith. Hierarchical 898 10,133 0.85 0.03 0.94 0.03 0.94 0.96 0.449 1.00 0.64 141,742 1.00 172
Pythia-6.9B Mod. Arith. Sparse probing 842 — 0.88 0.03 0.94 0.03 0.94 0.96 0.403 0.97 — 8,180 1.00 5
GPT-2 Large Parity Hierarchical 1 0 1.02 1.02 -0.05 1.02 -0.05 0.00 0.000 1.00 — 17 0.48 43 model at chance (0.48 vs 0.50): preservation is not meaningful; θ not reached
GPT-2 Medium Parity Hierarchical 1 0 1.00 0.99 0.29 1.00 0.29 -0.00 0.000 1.00 — 18 0.49 22 model at chance (0.49 vs 0.50): preservation is not meaningful
GPT-2 Small Parity Hierarchical 10 14 1.00 1.00 3.54 1.00 3.54 0.00 0.000 1.00 0.69 259 0.49 10 model at chance (0.49 vs 0.50): preservation is not meaningful
Llama-13B Parity Hierarchical — — — — — — — — — — — — 0.51 0 model at chance: no search
Llama-7B Parity Hierarchical 2 0 1.00 0.93 0.92 1.00 0.92 0.00 0.030 1.00 — 35 1.00 314
Pythia-1.4B Parity Hierarchical 767 6,973 0.98 0.57 0.95 0.87 0.95 0.40 0.104 1.00 0.77 94,737 0.95 144
Pythia-2.8B Parity Hierarchical 314 3,737 1.00 0.49 0.91 0.98 0.91 0.51 0.123 1.00 0.71 46,716 1.00 208
Pythia-6.9B Parity Hierarchical 336 3,568 0.94 0.50 0.88 0.70 0.88 0.49 0.177 1.00 0.57 45,439 1.00 419
GPT-2 Large Sorting Hierarchical 5,904 7,136 0.08 0.00 0.46 0.00 0.46 1.00 3.305 0.27 1.00 162,425 1.00 200 θ not reached
GPT-2 Medium Sorting Hierarchical 5,797 9,387 0.20 0.00 0.75 0.00 0.75 1.00 2.147 0.23 0.92 193,798 1.00 104 θ not reached
GPT-2 Small Sorting Hierarchical 5,164 6,309 0.78 0.00 0.96 0.00 0.96 0.99 0.205 0.43 1.00 139,939 0.99 48
Llama-13B Sorting Hierarchical 996 10,081 0.82 0.00 0.98 0.38 0.98 1.00 0.084 0.78 0.61 145,755 1.00 1,162
Llama-7B Sorting Hierarchical 1,391 8,703 0.83 0.00 0.99 0.00 0.99 1.00 0.064 0.74 0.92 135,065 1.00 724
Pythia-1.4B Sorting Hierarchical 3,524 32,983 0.84 0.00 0.97 0.00 0.97 1.00 0.121 0.23 0.77 497,152 1.00 135
Pythia-2.8B Sorting Hierarchical 2,227 20,812 0.85 0.00 0.97 0.00 0.97 1.00 0.148 0.23 0.66 313,891 1.00 201
Pythia-6.9B Sorting Hierarchical 2,272 8,818 0.85 0.00 0.98 0.03 0.98 1.00 0.142 0.21 1.00 156,343 1.00 377
GPT-2 Large WinoGrande Hierarchical 1 0 0.93 0.96 -0.01 0.87 -0.01 0.00 0.231 0.00 — 17 0.52 334 model at chance (0.52 vs 0.50): preservation is not meaningful; θ not reached
GPT-2 Medium WinoGrande Hierarchical 1 0 0.98 0.88 -0.51 0.89 -0.51 0.00 0.147 1.00 — 18 0.51 149 model at chance (0.51 vs 0.50): preservation is not meaningful; θ not reached
GPT-2 Small WinoGrande Hierarchical 40 32 1.01 1.02 0.42 1.02 0.42 0.01 0.155 0.07 1.00 891 0.48 79 model at chance (0.48 vs 0.50): preservation is not meaningful
Llama-13B WinoGrande Hierarchical 2,125 6,234 0.85 0.77 -0.14 0.76 -0.14 0.08 0.390 0.12 1.00 119,146 0.73 1,639 θ not reached
Llama-7B WinoGrande Hierarchical 2,232 4,732 0.90 0.74 0.62 0.77 0.62 0.10 0.229 0.12 1.00 97,214 0.68 1,207
Pythia-1.4B WinoGrande Hierarchical 1 0 0.96 0.98 -0.06 0.84 -0.06 0.02 0.171 1.00 — 18 0.55 191 model at chance (0.55 vs 0.50): preservation is not meaningful; θ not reached
Pythia-2.8B WinoGrande Hierarchical 1 0 0.86 0.87 -0.12 0.81 -0.12 0.00 0.223 0.00 — 17 0.61 217 θ not reached
Pythia-6.9B WinoGrande Hierarchical 1 0 0.88 0.88 -0.23 0.81 -0.23 0.00 0.251 0.00 — 18 0.60 450 θ not reached
Downloads last month
490

Paper for mudassiruddin/hagd