README.md

metadata

pretty_name: llmdbg analysis artifacts
license: mit
tags:
  - interpretability
  - mechanistic-interpretability
  - r-lens
  - jacobian-lens

llmdbg analysis artifacts

Versioned analysis data used by llmdbg.

Method references:

• Verbalizable Representations Form a Global Workspace in Language Models (J-Lens)
• R-Lens: Making J-Lens More Faithful on Early Layers

J-Lens

J-Lens artifacts contain averaged ordinary Jacobians from each residual layer to the final residual layer. They are fitted with Anthropic's reference jacobian-lens implementation, without the RelP backward-rule substitutions used for R-Lens.

R-Lens

R-Lens artifacts contain a learned residual-transport matrix for each source layer. Neuron Lens applies the appropriate transport to an MLP neuron's output direction, then applies the target model's final normalization and unembedding.

Each artifact directory contains:

• manifest.json: model compatibility, fitting provenance, tensor metadata, and SHA-256 integrity information.
• lens.safetensors: matrices named J.0, J.1, and so on.

The catalog includes locally fitted 0.8B and 2B Base maps. Their manifests explicitly allow compatible use with the corresponding Instruct checkpoints. That Base-to-Instruct reuse is a convenience approximation; an artifact fitted on the exact checkpoint is preferred for rigorous analysis.

The Qwen3.5 9B and 27B entries reference checkpoint-matched J-Lens and R-Lens maps from the R-Lens authors' pinned workspace-lenses release. The catalog records the immutable revision, file size, and SHA-256; llmdbg verifies the downloaded file before loading it. This avoids duplicating roughly 8.7 GB while keeping artifact resolution reproducible.