Hugging Face's logo

Datasets:
siddharthmb
/
2026.TA.tc8192_refusal_first10_graphs

Tags:
interpretability
circuit-tracing
attribution-graphs
transcoder-adapters
gemma-2-2b
Dataset card Files
xet
 Community
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:    CastError
Message:      Couldn't cast
auto_shard_gpus: bool
batch_size: int64
checkpoint: string
device: string
device_map: null
edge_threshold: double
features_dir: null
max_feature_nodes: int64
max_n_logits: int64
node_threshold: double
num_shards: int64
output_dir: string
port: int64
prompt_format: string
prompts: string
run_name: string
scan: string
serve: bool
shard_index: int64
graphs: list<item: struct<slug: string, scan: string, transcoder_list: list<item: null>, prompt_tokens: list (... 79 chars omitted)
  child 0, item: struct<slug: string, scan: string, transcoder_list: list<item: null>, prompt_tokens: list<item: stri (... 67 chars omitted)
      child 0, slug: string
      child 1, scan: string
      child 2, transcoder_list: list<item: null>
          child 0, item: null
      child 3, prompt_tokens: list<item: string>
          child 0, item: string
      child 4, prompt: string
      child 5, node_threshold: double
      child 6, schema_version: int64
to
{'graphs': List({'slug': Value('string'), 'scan': Value('string'), 'transcoder_list': List(Value('null')), 'prompt_tokens': List(Value('string')), 'prompt': Value('string'), 'node_threshold': Value('float64'), 'schema_version': 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
              auto_shard_gpus: bool
              batch_size: int64
              checkpoint: string
              device: string
              device_map: null
              edge_threshold: double
              features_dir: null
              max_feature_nodes: int64
              max_n_logits: int64
              node_threshold: double
              num_shards: int64
              output_dir: string
              port: int64
              prompt_format: string
              prompts: string
              run_name: string
              scan: string
              serve: bool
              shard_index: int64
              graphs: list<item: struct<slug: string, scan: string, transcoder_list: list<item: null>, prompt_tokens: list (... 79 chars omitted)
                child 0, item: struct<slug: string, scan: string, transcoder_list: list<item: null>, prompt_tokens: list<item: stri (... 67 chars omitted)
                    child 0, slug: string
                    child 1, scan: string
                    child 2, transcoder_list: list<item: null>
                        child 0, item: null
                    child 3, prompt_tokens: list<item: string>
                        child 0, item: string
                    child 4, prompt: string
                    child 5, node_threshold: double
                    child 6, schema_version: int64
              to
              {'graphs': List({'slug': Value('string'), 'scan': Value('string'), 'transcoder_list': List(Value('null')), 'prompt_tokens': List(Value('string')), 'prompt': Value('string'), 'node_threshold': Value('float64'), 'schema_version': 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.

2026.TA.tc8192_refusal_first10_graphs

Superseded (2026-10-14) by siddharthmb/2026.TA.tc8192_refusal_first10_headline_2026-10d_graphs: the same prompts rebuilt by the headline chain headline_2026-10d on the retrained adapter siddharthmb/2026.TA.gemma2_2b_headline_2026-10d_tc8192_decb_l1w0.001_auxk512_fp32_tarbb_lb2_ln1_sl17824146 with its feature store siddharthmb/2026.TA.features_2026.TA.gemma2_2b_headline_2026-10d_tc8192_decb_l1w0.001_auxk512_h47247ac74d15 (docs/decisions.md §9); this set stays as the record of the earlier adapter.

Base-vs-adapter circuit-tracing comparison overlay graphs for the tc8192_refusal_first10 run. Each overlay graph is two attribution graphs over the same prompt, the base model's (GemmaScope transcoder features, hexagon ⬢, with their reconstruction-error nodes, triangle ▲) and the adapter model's (adapter features, circle ●), built separately and drawn on identical tokens so the two circuits can be read position by position. Edges run within each side only: the sides share nothing but the embedding nodes, so no edge joins a base feature to an adapter feature, by construction. The fused kind (--graph_kind fused: one forward of the adapter model with every MLP replaced by T_base(x) + T_adapter(x) + Err, where base-adapter edges are real attribution) is where the two interact; a fused set is fused/ of siddharthmb/2026.TA.tc8192_refusal_first10_headline_2026-10d_graphs (docs/glossary.md Fused, overlay, relp).

What this run is

Where along the refusal opening does refusal commit? A base-vs-adapter overlay at each of the first ten tokens of the refusal opening — I / cannot / and / will / not / provide / instructions / for / how / <harm-word> — for 3 prompts (harm_125 meth, harm_139 DDoS, harm_116 pipe bomb) = 30 graphs. Deployed tc8192 adapter, native chat template.

Read the ADAPTER side: under chat the adapter is in-distribution and refusing (the focus), while the base side is the expected special-token gibberish. Step the dropdown "I → cannot → provide → …" to watch where the adapter's refusal/harmful features enter. This extends an earlier 2-position probe (positions "I" and " cannot", 3 prompts).

Companion results: a prefill-flip eval shows transplanting the instruct refusal opening onto BASE does not make it refuse; and diffing these graphs against the compliance build (2026.TA.tc8192_compliance_first10_graphs) shows 12/13 recurring "refusal-opening" features also fire when the adapter complies (generic response-start), leaving L6/F4241 as the only refusal-specific one (top-8 in ~33% of refusal graphs, 0 compliance graphs).

Findings / write-ups (repo): experiments/refusal_tracing/tex/refusal_token_tracing.pdf (full write-up + figures), experiments/refusal_tracing/results/graph_analysis.md, experiments/refusal_tracing/results/compare_refusal_compliance.md. Serve script: my_notes/launch_records/visualize_graphs/visualize 08-10-26 hf.sh (#3, port 8054).

How to use 

uv run --extra viz python -m analysis.serve siddharthmb/2026.TA.tc8192_refusal_first10_graphs:overlay,overlay_compact --port 8054

The index page lists the sets (overlay/, overlay_compact/) and the viewer opens each at /g/<set>/; name fewer sets after the colon to download less, and --compare <another repo or run dir> diffs this set against it graph by graph. The snapshot is downloaded and cached on first run (HF_HOME); node-click feature examples stream from the feature collections linked below.

Contents

path files size
overlay/ 31 8.6 GB
overlay_compact/ 31 21.6 MB
prompts/ 31 11.6 KB

base/ and adapter/, the one-sided graphs of the same run, were added to this repo by the disk audit of 2026-10-11 (claude_updates/memory/reports/disk_audit.md §7) before the scratch copy was deleted; the overlay graphs contain both sides.

Provenance

field value
adapter checkpoint siddharthmb/2026.TA.gemma2_2b_tc8192_decb_l1w0.001_tarbb_lb2.0_ln1_dr20000_lr8e-04_bs4_sl14793860
base model google/gemma-2-2b
adapter feature examples siddharthmb/2026.TA.features_2026.TA.gemma2_2b_tc8192_decb_l1w0.001_tarbb_lb2.0_ln1_dr20000_lr_he94e9602bafa
base feature examples siddharthmb/2026.TA.features_gemma-2-2b_gemmascope_width_16k_average_l0_76_ms100000_ml1024_tk2_h12ad59325ffd
prompt set /nlp/scr/siddharth/transcoder-adapters/base_adapter_comparisons/tc8192_refusal_first10/prompts
prompt format chat
base attribution targets union_adapter_top
GemmaScope google/gemma-scope-2b-pt-transcoders / width_16k / average_l0_76 (per-layer L0s in gemmascope_config.yaml)
overlay options {"hide_direct_embedding_logit_links": true, "compact_max_base_feature_nodes": 64, "compact_max_base_error_nodes": 0, "compact_max_adapter_feature_nodes": null}

Feature collections (dataset-side artifacts baked into the graphs):

Reproduce

Build invocation (from the run's Slurm log; the module name is updated to the current entry point, flags unchanged):

uv run --no-sync --extra viz python -m analysis.attribution.graphs.build_graphs --graph_kind overlay --adapter_checkpoint siddharthmb/2026.TA.gemma2_2b_tc8192_decb_l1w0.001_tarbb_lb2.0_ln1_dr20000_lr8e-04_bs4_sl14793860 --base_model google/gemma-2-2b --prompts /nlp/scr/siddharth/transcoder-adapters/base_adapter_comparisons/tc8192_refusal_first10/prompts --prompt_format chat --run_name tc8192_refusal_first10 --output_dir /nlp/scr/siddharth/transcoder-adapters/base_adapter_comparisons/tc8192_refusal_first10 --gemmascope_width width_16k --gemmascope_l0 average_l0_76 --gemmascope_l0_match nearest --feature_data_path siddharthmb/2026.TA.features_2026.TA.gemma2_2b_tc8192_decb_l1w0.001_tarbb_lb2.0_ln1_dr20000_lr_he94e9602bafa --base_feature_data_path siddharthmb/2026.TA.features_gemma-2-2b_gemmascope_width_16k_average_l0_76_ms100000_ml1024_tk2_h12ad59325ffd --continuation_max_new_tokens 0

Build log on the cluster: /juice2/u/siddharth/transcoder-adapters/logs/attribution/16717542.out

Re-star the jailbreak/compliance prompts in the dropdown after any rebuild (idempotent, auto-detects the set from the baked adapter continuations):

uv run python -m analysis.evals.classify_adapter_compliance \
  --graph_dir /nlp/scr/siddharth/transcoder-adapters/base_adapter_comparisons/tc8192_refusal_first10/overlay --star --also_star_dir /nlp/scr/siddharth/transcoder-adapters/base_adapter_comparisons/tc8192_refusal_first10/overlay_compact

Uploaded with (the module name is updated to the current entry point, flags unchanged):

uv run python -m analysis.attribution.hub --graph_dir /nlp/scr/siddharth/transcoder-adapters/base_adapter_comparisons/tc8192_refusal_first10

Cluster paths

  • graph dir: /nlp/scr/siddharth/transcoder-adapters/base_adapter_comparisons/tc8192_refusal_first10
  • manifest: /nlp/scr/siddharth/transcoder-adapters/base_adapter_comparisons/tc8192_refusal_first10/comparison-manifest.json (also in this repo)
  • build log: /juice2/u/siddharth/transcoder-adapters/logs/attribution/16717542.out
Downloads last month
2,604
Total file size:
14.2 GB