The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
metadata: struct<slug: string, scan: string, transcoder_list: list<item: null>, prompt_tokens: list<item: stri (... 519 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
child 7, comparison: struct<base_slug: string, adapter_slug: string, base_scan: string, adapter_scan: string, base_featur (... 338 chars omitted)
child 0, base_slug: string
child 1, adapter_slug: string
child 2, base_scan: string
child 3, adapter_scan: string
child 4, base_feature_scan: string
child 5, adapter_feature_scan: string
child 6, node_shapes: struct<base: string, adapter: string, shared: string>
child 0, base: string
child 1, adapter: string
child 2, shared: string
child 7, link_filter: struct<hide_direct_embedding_logit_links: bool, skipped_direct_embedding_logit_links: struct<base: i (... 22 chars omitted)
child 0, hide_direct_embedding_logit_links: bool
child 1, skipped_direct_embedding_logit_links: struct<base: int64, adapter: int64>
child 0, base: int64
child 1, adapter: int64
child 8, node_counts: struct<base_features: int64, adapter_features: int64, error_nodes: int64>
child 0, base_features: int64
...
d 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
child 7, comparison: struct<base_slug: string, adapter_slug: string, base_scan: string, adapter_scan: string, base_featur (... 250 chars omitted)
child 0, base_slug: string
child 1, adapter_slug: string
child 2, base_scan: string
child 3, adapter_scan: string
child 4, base_feature_scan: string
child 5, adapter_feature_scan: string
child 6, node_shapes: struct<base: string, adapter: string, shared: string>
child 0, base: string
child 1, adapter: string
child 2, shared: string
child 7, link_filter: struct<hide_direct_embedding_logit_links: bool, skipped_direct_embedding_logit_links: struct<base: i (... 22 chars omitted)
child 0, hide_direct_embedding_logit_links: bool
child 1, skipped_direct_embedding_logit_links: struct<base: int64, adapter: int64>
child 0, base: int64
child 1, adapter: int64
child 8, node_counts: struct<base_features: int64, adapter_features: int64, error_nodes: int64>
child 0, base_features: int64
child 1, adapter_features: int64
child 2, error_nodes: 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'), 'comparison': {'base_slug': Value('string'), 'adapter_slug': Value('string'), 'base_scan': Value('string'), 'adapter_scan': Value('string'), 'base_feature_scan': Value('string'), 'adapter_feature_scan': Value('string'), 'node_shapes': {'base': Value('string'), 'adapter': Value('string'), 'shared': Value('string')}, 'link_filter': {'hide_direct_embedding_logit_links': Value('bool'), 'skipped_direct_embedding_logit_links': {'base': Value('int64'), 'adapter': Value('int64')}}}, 'node_counts': {'base_features': Value('int64'), 'adapter_features': Value('int64'), 'error_nodes': 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 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 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
metadata: struct<slug: string, scan: string, transcoder_list: list<item: null>, prompt_tokens: list<item: stri (... 519 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
child 7, comparison: struct<base_slug: string, adapter_slug: string, base_scan: string, adapter_scan: string, base_featur (... 338 chars omitted)
child 0, base_slug: string
child 1, adapter_slug: string
child 2, base_scan: string
child 3, adapter_scan: string
child 4, base_feature_scan: string
child 5, adapter_feature_scan: string
child 6, node_shapes: struct<base: string, adapter: string, shared: string>
child 0, base: string
child 1, adapter: string
child 2, shared: string
child 7, link_filter: struct<hide_direct_embedding_logit_links: bool, skipped_direct_embedding_logit_links: struct<base: i (... 22 chars omitted)
child 0, hide_direct_embedding_logit_links: bool
child 1, skipped_direct_embedding_logit_links: struct<base: int64, adapter: int64>
child 0, base: int64
child 1, adapter: int64
child 8, node_counts: struct<base_features: int64, adapter_features: int64, error_nodes: int64>
child 0, base_features: int64
...
d 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
child 7, comparison: struct<base_slug: string, adapter_slug: string, base_scan: string, adapter_scan: string, base_featur (... 250 chars omitted)
child 0, base_slug: string
child 1, adapter_slug: string
child 2, base_scan: string
child 3, adapter_scan: string
child 4, base_feature_scan: string
child 5, adapter_feature_scan: string
child 6, node_shapes: struct<base: string, adapter: string, shared: string>
child 0, base: string
child 1, adapter: string
child 2, shared: string
child 7, link_filter: struct<hide_direct_embedding_logit_links: bool, skipped_direct_embedding_logit_links: struct<base: i (... 22 chars omitted)
child 0, hide_direct_embedding_logit_links: bool
child 1, skipped_direct_embedding_logit_links: struct<base: int64, adapter: int64>
child 0, base: int64
child 1, adapter: int64
child 8, node_counts: struct<base_features: int64, adapter_features: int64, error_nodes: int64>
child 0, base_features: int64
child 1, adapter_features: int64
child 2, error_nodes: 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'), 'comparison': {'base_slug': Value('string'), 'adapter_slug': Value('string'), 'base_scan': Value('string'), 'adapter_scan': Value('string'), 'base_feature_scan': Value('string'), 'adapter_feature_scan': Value('string'), 'node_shapes': {'base': Value('string'), 'adapter': Value('string'), 'shared': Value('string')}, 'link_filter': {'hide_direct_embedding_logit_links': Value('bool'), 'skipped_direct_embedding_logit_links': {'base': Value('int64'), 'adapter': Value('int64')}}}, 'node_counts': {'base_features': Value('int64'), 'adapter_features': Value('int64'), 'error_nodes': 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.
2026.TA.overlay_huge_graphs
Base-vs-adapter overlay attribution graphs (circuit-tracer JSON) for google/gemma-2-2b, contrasting the base GemmaScope transcoders against the huge 16384-feature transcoder adapter over the 40-prompt comprehensive set.
- base side — base GemmaScope transcoders (
width_16k,average_l0_76); reconstruct the base MLP, so this side stays base-like. - adapter side — the trained huge adapter
siddharthmb/2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl(2× capacity; B200 warm-start, final eval KL 0.142 / top-1 86.35%), which bridges base→instruct, so this side is instruct-like.
Why this repo exists (the fix). The original overlay_huge graph was built with run_combined_attribution, which runs ONE model on the base backbone — its forward collapses to base, so its completion/logit nodes were the base model's predictions (e.g. Answer/\n — base echoing/continuing the prompt), not the instruct-bridging adapter's. This repo is the corrected rebuild with run_base_adapter_comparison (runs the real adapter model + overlays a separate base run), so the adapter side shows instruction-tuned completions (e.g. Paris p≈0.96). Full story: my_notes/07-09-26/07-15-26 hybrid graph instruct-completions bug.md.
Contents
overlay/— 40 full base-vs-adapter overlay graphs (all node types).overlay_compact/— 40 compact overlays (error nodes stripped, base features capped) — the servable set.
Serve it
uv run --extra viz python -m analysis.attribution.serve_comparison_graphs \
--graph_file_dir siddharthmb/2026.TA.overlay_huge_graphs:overlay --port 8048
# or the compact set: ...:overlay_compact
Regenerate (corrected build)
uv run --extra viz python -m analysis.attribution.run_base_adapter_comparison \
--adapter_checkpoint siddharthmb/2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl \
--base_model google/gemma-2-2b --gemmascope_width width_16k --gemmascope_l0 average_l0_76 \
--base_feature_data_path siddharthmb/2026.TA.features_gemma-2-2b_gemmascope_width_16k_average_l0_76_ms100000_ml1024_tk1_hf83e96d574d2 \
--feature_data_path siddharthmb/2026.TA.features_2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2_hff15229f5fe1 \
--prompts analysis/attribution/prompts/comprehensive --prompt_format chat \
--max_feature_nodes 2048 --max_n_logits 10 \
--run_name overlay_huge_fixed --output_dir <out>
Built on a RunPod B200 (run_base_adapter_comparison, this session).
Inputs
- Base model:
google/gemma-2-2b; prompt tokenizergoogle/gemma-2-2b-it. - Huge adapter (adapter side):
siddharthmb/2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl. Training wandb:c7vfz8o0. - Base features:
siddharthmb/2026.TA.features_gemma-2-2b_gemmascope_width_16k_average_l0_76_ms100000_ml1024_tk1_hf83e96d574d2. - Adapter features:
siddharthmb/2026.TA.features_2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2_hff15229f5fe1.
Provenance
- Superseded original (base completions,
run_combined_attribution): local/nlp/scr/siddharth/transcoder-adapters/overlay_huge/graph. - Adapter training wandb:
siddharth-stanford/sparse-adaptation/runs/c7vfz8o0. - The attribution build itself is not wandb-logged.
- Downloads last month
- 36