josephofthebread/Qwen3-8B-concept-vectors

Behavioral concept vectors for Qwen/Qwen3-8B

1036 contrastive behavioral directions read off the residual stream of Qwen/Qwen3-8B, extracted from AntonKorznikov/feature_stories by difference of means over matched concept/antagonist story pairs.

Files

File	Method
diff.safetensors	mean(concept) - mean(antagonist)
concept_centered.safetensors	mean(concept) - mean(corpus)
antagonist_centered.safetensors	mean(antagonist) - mean(corpus)

We analyze layers 11, 14, 18, 22, and 25 of 36.

pairs.parquet has one row per pair in the same order as the pair axis, so row i names and scores tensor row i. It also carries the concept and antagonist labels, the ontology class, the narrative guidance, the story counts behind each pole, and per layer the norm and split-half reliability.

diff subtracts the antipole, which this corpus supports because both poles of a row describe the same premise, so topic, genre and language cancel exactly. The two *_centered tensors subtract the corpus mean. The following equality holds: diff == concept_centered - antagonist_centered.

Pooling is the mean over story tokens from position 50 onward. Each story is presented as an assistant turn behind a fixed, concept-free user prompt, with no system prompt.

Layers

Layer	Depth	Median Split-Half Cosine	Pairs Above 0.5	Median Diff Norm
11	0.3056	0.991	1036	3.3
14	0.3889	0.993	1036	4.4
18	0.5	0.995	1036	6.5
22	0.6111	0.995	1036	11.0
25	0.6944	0.995	1036	19.8

Geometry

Layer	Same Class	Different Class	Ratio	Anti-Aligned	Shared Component
11	11.3	0.398	0.246	0.237	1.61
14	10.7	0.410	0.256	0.244	1.60
18	10.0	0.432	0.268	0.252	1.61
22	10.7	0.434	0.261	0.244	1.66
25	11.0	0.425	0.253	0.240	1.68

Values are absolute cosines between the diff directions of two different pairs, split by whether the pairs belong to the same ontology class. For random unit vectors in 4096 dimensions the expected absolute cosine is 0.0125.