Hugging Face's logo

Datasets:
siddharthmb
/
2026.TA.hybrid_ft_overlay_ms100k_graphs

Tags:
circuit-tracer
attribution-graphs
interpretability
gemma-2-2b
transcoders
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:    TypeError
Message:      Couldn't cast array of type
struct<agree__cap_egypt: string, agree__cap_france: string, agree__cap_italy: string, agree__cap_japan: string, agree__continent_egypt: string, agree__first_president: string, agree__format_json_afterfence: string, agree__largest_ocean: string, agree__math_2plus2: string, agree__math_sqrt: string, agree__opposite_hot: string, agree__opposite_up: string, agree__planet_red: string, agree__speed_light: string, diverge__constraint_oneword_paris: string, diverge__explain_photosynthesis: string, diverge__format_json_brace: string, diverge__format_uppercase: string, diverge__haiku_autumn: string, diverge__json_only: string, diverge__list_colors: string, diverge__persona_shakespeare: string, diverge__polite_email: string, diverge__preamble_continue: string, diverge__recipe_req: string, diverge__steps_change_tire: string, diverge__summarize_req: string, diverge__sys_french_spain: string, interesting_small__bomb_refusal_help_I: string, interesting_small__capital_colesseum: string, interesting_small__capital_colesseum_mispelling: string, interesting_small__capital_paris: string, interesting_small__capital_paris_eos: string, interesting_small__cookies_explosives_however: string, interesting_small__favorite_drink: string, interesting_small__hacking_refusal_assist: string, interesting_small__math_forty_two_first: string, interesting_small__math_forty_two_second: string, interesting_small__opposite_cold: string, interesting_small__poison_refusal_help: string>
to
{'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')}
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 2303, in cast_table_to_schema
                  cast_array_to_feature(
                  ~~~~~~~~~~~~~~~~~~~~~^
                      table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                      feature,
                      ^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1852, in wrapper
                  return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
                                           ~~~~^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2059, in cast_array_to_feature
                  _c(array.field(name) if name in array_fields else null_array, subfeature)
                  ~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1854, in wrapper
                  return func(array, *args, **kwargs)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2149, in cast_array_to_feature
                  raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
              TypeError: Couldn't cast array of type
              struct<agree__cap_egypt: string, agree__cap_france: string, agree__cap_italy: string, agree__cap_japan: string, agree__continent_egypt: string, agree__first_president: string, agree__format_json_afterfence: string, agree__largest_ocean: string, agree__math_2plus2: string, agree__math_sqrt: string, agree__opposite_hot: string, agree__opposite_up: string, agree__planet_red: string, agree__speed_light: string, diverge__constraint_oneword_paris: string, diverge__explain_photosynthesis: string, diverge__format_json_brace: string, diverge__format_uppercase: string, diverge__haiku_autumn: string, diverge__json_only: string, diverge__list_colors: string, diverge__persona_shakespeare: string, diverge__polite_email: string, diverge__preamble_continue: string, diverge__recipe_req: string, diverge__steps_change_tire: string, diverge__summarize_req: string, diverge__sys_french_spain: string, interesting_small__bomb_refusal_help_I: string, interesting_small__capital_colesseum: string, interesting_small__capital_colesseum_mispelling: string, interesting_small__capital_paris: string, interesting_small__capital_paris_eos: string, interesting_small__cookies_explosives_however: string, interesting_small__favorite_drink: string, interesting_small__hacking_refusal_assist: string, interesting_small__math_forty_two_first: string, interesting_small__math_forty_two_second: string, interesting_small__opposite_cold: string, interesting_small__poison_refusal_help: string>
              to
              {'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')}

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_overlay_ms100k_graphs

Base-vs-adapter overlay attribution graphs (circuit-tracer JSON) for google/gemma-2-2b, built with the corrected hybrid pipeline. Each graph overlays two runs on the same prompt:

On Answer with one word. What city is the Eiffel Tower in? … model\nParis, the base side's top next-token logit is '\n' (p=0.617, continue-document) while the adapter side's is ' ' (p=0.902, stop/answer) — the two behaviors this overlay is meant to contrast.

Why this repo exists: an earlier hybrid graph built with run_combined_attribution showed base completions on both sides (its forward collapses to the base backbone). The fix was to rebuild with run_base_adapter_comparison (overlay of a real adapter run + a base run). Full story: my_notes/07-15-26 hybrid graph instruct-completions bug.md in the [transcoder-adapters repo].

Contents

  • overlay_compact/ — 12 compact (base64, error nodes stripped) overlay graphs + graph-metadata.json. This is what you serve.
  • comparison-manifest.json — the build manifest (models, prompts, per-prompt graph hashes, output paths).

These are not a tabular dataset (no experiment-name column) — they are per-prompt circuit-tracer graph JSONs. The equivalent identifier is the run_name hybrid_ft_overlay_ms100k plus the prompt slug in each filename.

Prompts (file → prompt intent)

prompt slug intent
capital_paris / capital_paris_eos Eiffel Tower → Paris (with/without EOS), base-vs-instruct completion contrast
capital_colesseum / capital_colesseum_mispelling Colosseum → Rome (clean / misspelled)
favorite_drink open-ended preference
opposite_cold antonym (cold → hot)
math_forty_two_first / math_forty_two_second arithmetic to 42
bomb_refusal_help_I / poison_refusal_help / hacking_refusal_assist / cookies_explosives_however refusal / safety-relevant prompts

Serve it

Requires the transcoder-adapters repo (the visualizer resolves this repo id via snapshot_download):

uv run --extra viz python -m analysis.attribution.serve_comparison_graphs \
  --graph_file_dir siddharthmb/2026.TA.hybrid_ft_overlay_ms100k_graphs:overlay_compact --port 8046

Regenerate 

uv run --extra viz python -m analysis.attribution.run_base_adapter_comparison \
  --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 --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 siddharthmb/2026.TA.features_gemma-2-2b_gemmascope_width_16k_average_l0_76_ms100000_ml1024_tk2_h12ad59325ffd \
  --feature_data_path siddharthmb/2026.TA.features_2026.TA.gemma2_2b_tc8192_decb_l1w0.001_tarbb_lb2.0_ln1_dr20000_lr_he94e9602bafa \
  --prompts analysis/attribution/prompts/interesting_small --prompt_format chat \
  --max_feature_nodes 2048 --max_n_logits 10 \
  --run_name hybrid_ft_overlay_ms100k \
  --output_dir $LARGE_ARTIFACTS_DIR/transcoder-adapters/base_adapter_comparisons/hybrid_ft_overlay_ms100k

Inputs

Provenance / cluster paths

  • Build job: Slurm 16194543 (COMPLETED, 3m39s, jagupard35), run_name=hybrid_ft_overlay_ms100k.
  • Logs: logs/attribution/20260715_190509_16194543.out / .err in the transcoder-adapters repo.
  • Source dir on cluster: /nlp/scr/siddharth/transcoder-adapters/base_adapter_comparisons/hybrid_ft_overlay_ms100k/ (sides base/, adapter/, overlay/, overlay_compact/).
  • wandb: none (attribution graph builds are not wandb-logged).

comprehensive/ — 40-prompt set across categories + 3-model completions

A larger overlay set (run_name=hybrid_ft_overlay_ms100k_comprehensive, Slurm 16198499) over analysis/attribution/prompts/comprehensive (40 prompts), plus greedy 12-token completions from three models for every prompt.

  • comprehensive/overlay_compact/ — 40 compact overlay graphs. Serve: --graph_file_dir siddharthmb/2026.TA.hybrid_ft_overlay_ms100k_graphs:comprehensive/overlay_compact.
  • comprehensive/completions.json / .md — for each prompt, the 12-token greedy continuation from base google/gemma-2-2b, our hybrid (adapter model), and instruct google/gemma-2-2b-it, all given the identical gemma-2-2b-it chat-formatted input.

Prompt categories: agree__ (14, base and instruct agree — factual/math/antonym controls), diverge__ (14, base and instruct diverge — format constraints, personas, instruction-following), interesting_small__ (12, refusal/safety + capitals). On the diverge__ prompts the completions show base continuing/echoing the document while the hybrid follows the instruction, closely tracking instruct — direct evidence the adapter carries instruction-following.

interesting_queries/ — 48-prompt refusal/jailbreak set + 3-model completions

Overlay (run_name=hybrid_ft_overlay_ms100k_interesting_queries, Slurm 16198944) over analysis/attribution/prompts/interesting_queries (48 prompts) plus greedy 12-token completions from base/hybrid/instruct.

  • interesting_queries/overlay_compact/ — 42 compact overlay graphs (the 6 adv_suffix/*_no_suffix files are byte-identical to their harmful/harm_0XX counterparts, so they collapse by prompt-token hash). Serve: --graph_file_dir siddharthmb/2026.TA.hybrid_ft_overlay_ms100k_graphs:interesting_queries/overlay_compact.
  • interesting_queries/completions.json / .md — base/hybrid/instruct 12-token continuations.

Categories: harmful/ (18, harm_XXX harmful requests), divergent/ (18, div_XXXX base/instruct-divergent), adv_suffix/ (12, 6 harmful queries each with/without an adversarial jailbreak suffix). Finding: on harmful prompts the hybrid refuses like instruct ("I cannot provide instructions…") while base complies/echoes — refusal behavior lives in the adapter. On the adv_suffix variants the suffix nudges instruct toward engaging while the hybrid still refuses.
Downloads last month
23
Total file size:
89.8 MB