Downloads · 30 days
0
neil0306/jlensvl-qwen35-4b-lens
jlensvl-qwen35-4b-lens is a machine learning model from neil0306. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
A fitted J-Lens (Jℓ) for Qwen/Qwen3.5-4B, produced with JLensVL, a vision-language extension of Anthropic's jacobian-lens reference engine. The lens transports a residual-stream activation into the final-layer basis w…
Downloads · 30 days
0
Access
Public
Updated Jul 11, 2026
Repo size
813 MB
Likes
1
Public
Click a slice to open those files.
.pt406 MB · 50%
From the Hugging Face model README
A fitted J-Lens (J_ℓ) for Qwen/Qwen3.5-4B, produced with
JLensVL, a vision-language extension of Anthropic's
jacobian-lens reference engine. The lens transports a
residual-stream activation into the final-layer basis with the model's average input→output Jacobian
and decodes it with the unembedding — a much sharper readout than the plain logit-lens, especially
mid-stack, where the logit-lens is mostly noise.
lens_qwen35_4b_final.pt — the fitted lens (torch), loadable via JacobianLens.from_pretrained(...)
or JLensVL.from_pretrained(model_id, lens=...).lens_jl.npz — the same J_ℓ matrices exported to a plain .npz, for the MLX forward-only backend
on Apple Silicon (jlensvl.mlx_backend.MLXJLens) — no custom Metal backward kernel needed.from jlensvl import JLensVL
jl = JLensVL.from_pretrained("Qwen/Qwen3.5-4B", lens="lens_qwen35_4b_final.pt")
print(jl.trace("Fact: The currency used in the country shaped like a boot is the")[30])
# -> euro / Italian
# Apple Silicon, forward-only
from jlensvl.mlx_backend import MLXJLens
jl = MLXJLens.from_pretrained("mlx-community/Qwen3.5-4B-4bit", "lens_jl.npz")
print(jl.trace("... the country shaped like a boot is the")[30])
See the JLensVL README for the full API (vision tracing, concept-race conflict analysis, and the prompt helper).
Fitted on an Apple M4 Max (torch + MPS backend, bf16) on a small curated prompt corpus. Not the
fla/causal-conv1d fast path — that must stay uninstalled for the GDN layers to remain
differentiable.