README.md

metadata

license: cc-by-4.0
language:
  - en
  - de
  - fr
  - zh
tags:
  - interpretability
  - llm-channels
  - wordnet
pretty_name: LCN Paper Data (TPAMI channels paper)

LCN Paper Data

Key data backing the figures and lexical analysis of the TPAMI channels paper.

dictionaries/ (词典 / 大全)

Token -> WordNet sense dictionaries.

• super_thesaurus.json (2.1 MB): flat dict {word: "<wordnet.lexname>"} (词典).
• expanded_thesaurus.json (6.5 MB): {word: ["sense.n.01", ...]} full senses (大全).

bridge_coords/

2-D PCA coords for the fig:bridge panels (pooled + per-token).

• phase3a_*: cross-lingual EN vs DE/FR/ZH (Llama3-1B, Llama3-8B, Qwen3-1.7B, Qwen3-8B).
• phase3b_*: cross-modal text vs vision-language (Qwen2.5-VL-7B, Qwen3-VL-2B, Qwen3-VL-8B, InternVL3-8B). The *_pertoken.json files carry a top_k list so an alternate token can be picked per panel with no GPU rerun.

FIGURE_SOURCE_DATA_MAP.md

Catalog: each figure -> raw data file -> server -> generation script. GB-scale raw direction_space_raw.json live on the Leonardo cluster and are not mirrored here.

token_channel_dicts/ (token <-> feature/channel-dimension 大词典)

Maps each shared-concept token to the model feature dimensions (channels) it activates. This is the token<->channel mapping behind the figures.

multilingual_llm/fig10_ml_shared_channels.json (2.9 MB)

{ "shared_channels": { "": { "shared_channels": ["individual_18_K_18_ch937", ...], "n_shared": int, "overlap_ratio": float, "pair_overlaps": ..., "per_lang_top": ... } }, "pca": {...} } Channel ID format encodes the feature/channel index C.

multimodal_vlm/ (4 VLMs: internvl3_2b, internvl3_8b, qwen3vl_2b, llava_ov_7b)

• mm_shared_token_channels_.json: { items: [ { token, channels: [ {layer, channel, max_activation}, ... ] } ] } -- channels where the token is top-10 in BOTH vision-language and text.
• mm_bridge_per_token_channels_.json: cross-modal bridge, per-token top-k channels per layer (layers 15/20/25). Here is the literal feature-dimension index within .

token_channel_dicts/ (token <-> feature/channel-dimension big dictionary)

Maps each shared-concept token to the model feature dimensions (channels) it activates. This is the token<->channel mapping behind the figures.

multilingual_llm/fig10_ml_shared_channels.json (2.9 MB)

Schema: {"shared_channels": {"<token>": {"shared_channels": ["individual_18_K_18_ch937", ...], "n_shared": int, "overlap_ratio": float, "pair_overlaps": ..., "per_lang_top": ...}}, "pca": {...}} The channel ID string individual_<i>_K_<k>_ch<C> encodes the feature/channel index C.

multimodal_vlm/ (4 VLMs: internvl3_2b, internvl3_8b, qwen3vl_2b, llava_ov_7b)

• mm_shared_token_channels_<model>.json: {"items": [{"token": str, "channels": [{"layer": str, "channel": int, "max_activation": float}, ...]}]} channels where the token is top-10 in BOTH vision-language and text.
• mm_bridge_per_token_channels_<model>.json: cross-modal bridge, per-token top-k channels per layer (layers 15/20/25). Here channel is the literal feature-dimension index within layer.

thesaurus/ (channel self-aggregation: grouping + selectivity)

Training-free analysis of whether a channel's top-activating tokens cluster into near-synonyms / same-category words (indicator 1), and whether a channel monopolizes specific dataset tokens (indicator 3). 6 models: Llama-3.2-1B, Llama-3.1-8B, Gemma-3-1B, Gemma-3-4B, Qwen3-1.7B, Qwen3-8B (self-gen on tatsu-lab/alpaca). Uses the dictionaries/ files above.

Top-level tables (self-contained)

• MASTER_TABLE.md - master table: 6 models x K in {3,5,10,20,30,50} x 5 grouping metrics x 3 methods (WordNet-synonym / WordNet-category / MiniLM cos>=0.4/0.45/0.5) x 2 scopes, real vs random, + indicator 3.
• SELECTIVITY_SWEEP_TABLE.md - selectivity@K real vs random, 6 models.
• SELECTIVITY_SCALE_COMPARE.md - 400 -> 5000/20000 data-scaling de-bias.
• SELECTIVITY_ACT_TABLE.md - top-1-by-activation selectivity (noise-diagnosed).

Indicator 3 selectivity

selectivity@K(channel) = mean of min(num/den,1) over the channel's top-K activation tokens. num = how often the channel high-activates that token (na-count); den = that token's total occurrences in all N self-gen answers (tokenizer, strip-aligned). Random = redistribution null (redistribute each token's high-activation events uniformly across the module's channels) -> 0 at scale. Finding: scaling data removes small-sample inflation, but real stays >> random (10x at 20k). qwen3_1b has full 20000 (data/sel_sweep_qwen3_1b_n20000.json); others at 400 + partial scale.

data/ and code/

• data/sel_sweep_<model>[_na400|_n5000|_n20000].json - selectivity@K per sample size.
• data/random_v2_* (vocab/alpaca null), data/random_perm_* (permutation null), data/results_*, data/sel_act_*.
• code/ - grouping core (thesaurus_grouping.py), selectivity_sweep/act, group_metric_sweep, gen_master_table, random_baseline_v2, split/merge_chunks (20k chunk->merge pipeline), build_super_thesaurus.