Hugging Face's logo

Datasets:
siddharthmb
/
2026.TA.hybrid_ft_ftexamples_graphs

ArXiv:
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
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
base_feature_scan: string
n_base_features: int64
output_dir: string
run_name: string
results: struct<bomb_refusal_help_I: string, capital_colesseum: string, capital_colesseum_mispelling: string, (... 255 chars omitted)
  child 0, bomb_refusal_help_I: string
  child 1, capital_colesseum: string
  child 2, capital_colesseum_mispelling: string
  child 3, capital_paris: string
  child 4, capital_paris_eos: string
  child 5, cookies_explosives_however: string
  child 6, favorite_drink: string
  child 7, hacking_refusal_assist: string
  child 8, math_forty_two_first: string
  child 9, math_forty_two_second: string
  child 10, opposite_cold: string
  child 11, poison_refusal_help: string
adapter_checkpoint: string
base_model: string
max_error_nodes: int64
adapter_feature_scan: string
to
{'run_name': Value('string'), 'adapter_checkpoint': Value('string'), 'base_model': Value('string'), 'n_base_features': Value('int64'), 'output_dir': Value('string'), 'base_feature_scan': Value('string'), 'adapter_feature_scan': Value('string'), 'max_error_nodes': Value('int64'), 'results': {'bomb_refusal_help_I': Value('string'), 'capital_colesseum': Value('string'), 'capital_colesseum_mispelling': Value('string'), 'capital_paris': Value('string'), 'capital_paris_eos': Value('string'), 'cookies_explosives_however': Value('string'), 'favorite_drink': Value('string'), 'hacking_refusal_assist': Value('string'), 'math_forty_two_first': Value('string'), 'math_forty_two_second': Value('string'), 'opposite_cold': Value('string'), 'poison_refusal_help': Value('string')}}
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
              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
              base_feature_scan: string
              n_base_features: int64
              output_dir: string
              run_name: string
              results: struct<bomb_refusal_help_I: string, capital_colesseum: string, capital_colesseum_mispelling: string, (... 255 chars omitted)
                child 0, bomb_refusal_help_I: string
                child 1, capital_colesseum: string
                child 2, capital_colesseum_mispelling: string
                child 3, capital_paris: string
                child 4, capital_paris_eos: string
                child 5, cookies_explosives_however: string
                child 6, favorite_drink: string
                child 7, hacking_refusal_assist: string
                child 8, math_forty_two_first: string
                child 9, math_forty_two_second: string
                child 10, opposite_cold: string
                child 11, poison_refusal_help: string
              adapter_checkpoint: string
              base_model: string
              max_error_nodes: int64
              adapter_feature_scan: string
              to
              {'run_name': Value('string'), 'adapter_checkpoint': Value('string'), 'base_model': Value('string'), 'n_base_features': Value('int64'), 'output_dir': Value('string'), 'base_feature_scan': Value('string'), 'adapter_feature_scan': Value('string'), 'max_error_nodes': Value('int64'), 'results': {'bomb_refusal_help_I': Value('string'), 'capital_colesseum': Value('string'), 'capital_colesseum_mispelling': Value('string'), 'capital_paris': Value('string'), 'capital_paris_eos': Value('string'), 'cookies_explosives_however': Value('string'), 'favorite_drink': Value('string'), 'hacking_refusal_assist': Value('string'), 'math_forty_two_first': Value('string'), 'math_forty_two_second': Value('string'), 'opposite_cold': Value('string'), 'poison_refusal_help': Value('string')}}
              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.hybrid_ft_ftexamples_graphs

Base-vs-adapter attribution graphs for the hybrid_ft_ftexamples run, built with python -m analysis.attribution.graphs.build_graphs --graph_kind fused with graph kind(s) fused; the graphs sit at the repo root.

Combined base+adapter attribution graphs for the hybrid_ft_ftexamples run: ONE attribution graph per prompt over a single forward pass of the model named by the backbone below. Each MLP is replaced by T_base(x) + T_adapter(x) + Err, so every graph holds GemmaScope base features (hexagon ⬢), transcoder-adapter features (circle ●) and reconstruction-error nodes (triangle ▲) under one shared set of logit nodes; base<->adapter edges and edges into the logits are real attribution. This differs from the overlay graphs (--graph_kind overlay), which glue two separately computed graphs that share only embedding nodes.

Backbone: base. The combined replacement runs on the plain google/gemma-2-2b backbone, whose forward pass collapses to the base model, so the logits and continuations are the BASE model's; read adapter features here as a decomposition of the base model's computation, not as the adapter model's circuit. (The manifest predates the backbone key; all such runs used the base backbone.)

A 2026-07-15/16 overlay set of the input-shift study (experiments/transcoder_input_shift): the fine-tuned GemmaScope layers L0, L24, L25 on the instruct side (the hybrid set) against the base GemmaScope, with the completions baked beside the graphs.

Uploaded unchanged from the cluster's scratch volume by the disk audit of 2026-10-11 (claude_updates/memory/reports/disk_audit.md) so the scratch copy could be deleted. A historical set: the refusal-tracing, fused-graph and input-shift writeups cite the sets they name, and the headline rerun of 2026-10 (siddharthmb/2026.TA.gemma2_2b_headline_2026-10d_tc8192_decb_l1w0.001_auxk512_fp32_tarbb_lb2_ln1_sl17824146; the sets siddharthmb/2026.TA.tc8192_refusal_first10_headline_2026-10d_graphs, siddharthmb/2026.TA.tc8192_compliance_first10_headline_2026-10d_graphs) supersedes the adapter-side claims (docs/decisions.md §9, §11).

How to use 

uv run --extra viz python -m analysis.serve siddharthmb/2026.TA.hybrid_ft_ftexamples_graphs --port 8053

The graphs sit at the repo root, so the bare repo id is the graph dir (no :subdir). The snapshot is downloaded and cached on first run (HF_HOME); node-click feature examples stream from the feature collections linked below.

How it was produced

Produced by python -m analysis.attribution.graphs.build_graphs --graph_kind fused from transcoder-adapters.

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

uv run --no-sync python -m analysis.attribution.graphs.build_graphs --graph_kind fused --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 analysis/attribution/prompts/interesting_small --prompt_format chat --gemmascope_width width_16k --gemmascope_l0 average_l0_76 --finetuned_transcoder_dir siddharthmb/2026.TA.gemma2_2b_gemmascope_transcoders_instruct_ft_L0-24-25 --finetuned_layers 0 24 25 --base_feature_data_path /nlp/scr/siddharth/transcoder-adapters/feature_data/base_ftL0-24-25_merge20 --adapter_feature_data_path siddharthmb/2026.TA.features_2026.TA.gemma2_2b_tc8192_decb_l1w0.001_tarbb_lb2.0_ln1_dr20000_lr_he94e9602bafa --max_feature_nodes 4096 --batch_size 4 --max_n_logits 5 --max_error_nodes 32 --run_name hybrid_ft_ftexamples --output_dir /nlp/scr/siddharth/transcoder-adapters/base_adapter_comparisons/hybrid_ft_ftexamples

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 /juice2/scr2/siddharth/transcoder-adapters/base_adapter_comparisons/hybrid_ft_ftexamples --star

Uploaded with:

uv run python -m analysis.attribution.hub --graph_dir /juice2/scr2/siddharth/transcoder-adapters/base_adapter_comparisons/hybrid_ft_ftexamples

Details

Contents

path files size
graphs (repo root) 12 285.3 MB
top-level files 2 10.4 KB

Inputs

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 /base_features
prompt set ``
prompt format ``
backbone base (manifest predates the key)
n_base_features 16384 (combined feature index < this is a base feature)
max_error_nodes 32
prompts traced 12 of 12 succeeded

Results

Graphs

graph base features adapter features error nodes
hybrid_ft_ftexamples__bomb_refusal_help_I__hafcdd5f8d6c1 n/a n/a n/a
hybrid_ft_ftexamples__capital_colesseum__he6b0cdbc003d n/a n/a n/a
hybrid_ft_ftexamples__capital_colesseum_mispelling__h6c3a39bbc532 n/a n/a n/a
hybrid_ft_ftexamples__capital_paris__h4f93ff3e59c4 n/a n/a n/a
hybrid_ft_ftexamples__capital_paris_eos__h0778de7116be n/a n/a n/a
hybrid_ft_ftexamples__cookies_explosives_however__h7d5edf13528e n/a n/a n/a
hybrid_ft_ftexamples__favorite_drink__h52755272bf00 n/a n/a n/a
hybrid_ft_ftexamples__hacking_refusal_assist__h56595634e1ac n/a n/a n/a
hybrid_ft_ftexamples__math_forty_two_first__h1426d9baeaab n/a n/a n/a
hybrid_ft_ftexamples__math_forty_two_second__h4bf177829657 n/a n/a n/a
hybrid_ft_ftexamples__opposite_cold__h407abc63d121 n/a n/a n/a
hybrid_ft_ftexamples__poison_refusal_help__h04f174d68262 n/a n/a n/a

Node counts are n/a because this run's graph-metadata.json predates recording them; each graph's metadata.comparison.node_counts has them.

Links

Cluster paths

  • graph dir: /juice2/scr2/siddharth/transcoder-adapters/base_adapter_comparisons/hybrid_ft_ftexamples
  • manifest: /juice2/scr2/siddharth/transcoder-adapters/base_adapter_comparisons/hybrid_ft_ftexamples/combined_attribution_manifest.json (also in this repo)
  • build log: /juice2/u/siddharth/transcoder-adapters/logs/combined_attribution/20260715_222426_16196695.out
Downloads last month
73
Total file size:
299 MB

Paper for siddharthmb/2026.TA.hybrid_ft_ftexamples_graphs