2026.TA.features_gemma-2-2b_gemmascope_width_16k_average_l0_76_ms5000_ml1024_tk10_h8f13c4bce9be

2026.TA.features_gemma-2-2b_gemmascope_width_16k_average_l0_76_ms5000_ml1024_tk10_h8f13c4bce9be

Packed circuit-tracer feature cache produced by analysis.features.collect.

Status

Complete.

How to use

Attribution graphs read this repo as a feature scan: pass --feature_data_path siddharthmb/2026.TA.features_gemma-2-2b_gemmascope_width_16k_average_l0_76_ms5000_ml1024_tk10_h8f13c4bce9be (adapter features) or --base_feature_data_path siddharthmb/2026.TA.features_gemma-2-2b_gemmascope_width_16k_average_l0_76_ms5000_ml1024_tk10_h8f13c4bce9be (GemmaScope base features) to python -m analysis.attribution.graphs.build_graphs; the viewer started by python -m analysis.attribution.serve then streams a clicked node's examples from features/index.json.gz and features/layer_N.bin in this repo.

Read one feature by its cantor id without the viewer:

from analysis.features.collect.hub import load_circuit_tracer_feature_from_hub
from helpers.ids import cantor_pair

feature = load_circuit_tracer_feature_from_hub("siddharthmb/2026.TA.features_gemma-2-2b_gemmascope_width_16k_average_l0_76_ms5000_ml1024_tk10_h8f13c4bce9be", cantor_pair(layer, idx))

Browse the whole collection with the dashboard, which reads a collection dir on the cluster (feature_metadata.json, the packed store circuit_tracer_features/, the histogram and annotation sidecars):

uv run python -m analysis.features.dashboard --data_dir <the collection's output dir on the cluster>

Collection Inputs

• Model: gemma-2-2b_gemmascope_width_16k_average_l0_76
• Config: feature_collection_config.json
• Datasets:
  • siddharthmb/2026.transcoder-adapters.lmsys-chat-1m-splits
  • science-of-finetuning/fineweb-1m-sample

Summary

• Total tokens: 4424834
• Tokens per domain: {'chat': 1979712, 'fineweb': 2445122}