You can deploy this model and start earning money today!

agamemnon/
talkie-1930-13b-base-jspace

Name: talkie-1930-13b-base-jspace
Availability: InStock

jlens

jacobian-lens

jspace

mechanistic-interpretability

Views

No views yet

Talkie 1930 13B Base J-Space

This repository contains a Jacobian-lens J-space matrix for xlr8harder/talkie-1930-13b-base-tf.

Freshly fit by this project

This matrix was freshly fit by the Talkie J-Space project. It was computed from the pinned Talkie checkpoint and shared prompt corpus recorded in manifest.json. It was not copied, converted, averaged, or downloaded from any pre-existing Talkie J-space or Jacobian-lens artifact.

The repository intentionally contains no Talkie model weights and no fitting checkpoint.

Matrix and fit

Field	Value
Model	`xlr8harder/talkie-1930-13b-base-tf`
Model revision	`7cf020702ef7fd68ac16c85aaea389206ded519e`
Source layers	0–38
Target layer	39
Residual width	5120
Prompts completed	100
Maximum prompt length	128 tokens
Dimension batch	56
GPU	NVIDIA RTX PRO 6000 Blackwell Server Edition
Corpus	`Salesforce/wikitext`
Corpus revision	`b08601e04326c79dfdd32d625aee71d232d685c3`
Jacobian Lens revision	`581d398613e5602a5af361e1c34d3a92ea82ba8e`
Artifact SHA-256	`21dbccaedd412451e1b8819d192d98af5c792d438d0d228f2f5b3f8d5afbfff9`
Artifact bytes	2044734799

The artifact has 39 fp16 matrices of shape 5120 × 5120, mapping residual states from layers 0–38 into the layer-39 basis. manifest.json is the machine-readable source of truth for exact provenance and checksum verification.

Loading

python

from jlens import JacobianLens

lens = JacobianLens.from_pretrained("agamemnon/talkie-1930-13b-base-jspace")
print(lens.source_layers, lens.d_model)

Use the exact model revision in manifest.json with the Talkie adapter from the Talkie J-Space backend. A Jacobian lens is architecture- and revision-specific.

Limitations

This is an interpretability research artifact, not a factuality or safety classifier. J-space readouts can be misleading and should be validated against model behavior. Raw base-model output can be inaccurate, biased, or unsuitable for all audiences.

License and attribution

The matrix and project-authored metadata are released under Apache-2.0. Upstream Apache-2.0, MIT, CC-BY-SA, and GFDL attributions are in THIRD_PARTY_NOTICES.md. The underlying model remains subject to its own model card and license.