AxesLens

Ready-to-use geometric axes for all 1,999 WordNet semantic axes (antonymous adjective synset pairs, e.g. poor vs. rich), recovered from every attention head and layer of Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.1 and Qwen3-8B. Use them to rate concepts on semantic axes, find the heads that encode an axis, or steer generation. Code and paper: https://github.com/NsrxRwWa/AxesLens

Files

<model>/<template>_n<n>/<level>_<position>.npz

• model: Meta-Llama-3-8B-Instruct | Mistral-7B-Instruct-v0.1 | Qwen3-8B
• template, n: listing | simple; 15 | 30 sentences per pole
• level: head (attention heads) | layer (residual stream)
• position: mean (mean over tokens) | last (last token)

Each file contains axis_keys (semantic axes), scores (variance ratios) and directions (geometric axes). Shapes:

Model	head	layer
Llama-3-8B / Mistral-7B	(1999, 32, 32, 128)	(1999, 32, 1, 4096)
Qwen3-8B	(1999, 36, 32, 128)	(1999, 36, 1, 4096)

Shapes are for directions (axes x layers x heads x dims); scores drops the last dimension.

Usage

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download("NsrxRwWa/AxesLens", "Meta-Llama-3-8B-Instruct/listing_n30/head_mean.npz",
                       repo_type="dataset")
d = np.load(path)
keys, scores, directions = d["axis_keys"], d["scores"], d["directions"]

# Top-10 heads (layer, head) of semantic axis i, ranked by variance ratio
i = 0
top = np.argsort(scores[i].ravel())[::-1][:10]
layers, heads = np.unravel_index(top, scores.shape[1:])
print(keys[i], list(zip(layers, heads)))

theta = directions[i, layers, heads]      # geometric axes of these heads

Intended use and risks

Intended for research on how LLMs represent concepts, e.g. auditing and mitigating stereotypical associations. Out of scope: using the geometric axes to steer models toward stereotypical, discriminatory, or harmful outputs, or to make decisions about individuals or groups. Projections reflect associations encoded in the models, not facts about social groups. Users must also comply with the Meta Llama 3 Community License (including its Acceptable Use Policy), the Mistral license and the Qwen3 license (Apache 2.0).