bluelightai/clt-qwen3-0.6b-base-20k

Cross-layer transcoder for Qwen3-0.6B-Base

Blog | Technical Report | Feature Dashboard | BluelightAI

This is a cross-layer transcoder trained to interpret the activations of Qwen3-0.6B-Base. It can be used with the open source circuit-tracer library to build and interrogate attribution graphs for prompts. You can also explore the features on our dashboard.

What is a cross-layer transcoder?

A cross-layer transcoder is an interpreter model trained to extract sparsely-activating features from the activations of a Transformer model. Its encoder translates the input $x_{\mathrm{\ell }}^{\text{in}}x^\{\backslash text\{in\}\}\_\{\backslash ell\}$ of each MLP layer of the Transformer into a high-dimensional but sparse feature vector $f_{\mathrm{\ell }}f\_\backslash ell$. The decoder then reconstructs the output $x_{\mathrm{\ell }}^{\text{out}}x^\{\backslash text\{out\}\}\_\{\backslash ell\}$ of the MLP using features that were extracted from all previous layers. In formulas:

$$f_{\mathrm{\ell }}=\sigma (W_{\mathrm{\ell }}^{\text{enc}}{x}_{\mathrm{\ell }}^{\text{in}}+b_{\mathrm{\ell }}^{\text{enc}})f\_\backslash ell\; =\; \backslash sigma(W^\{\backslash text\{enc\}\}\_\backslash ell\; x^\{\backslash text\{in\}\}\_\{\backslash ell\}\; +\; b^\{\backslash text\{enc\}\}\_\backslash ell)$$ $${\widehat{x}}_{\mathrm{\ell }}^{\text{out}}=\sum _{k\le \mathrm{\ell }}{W}_{k\to \mathrm{\ell }}^{\text{dec}}{f}_k+b_{\mathrm{\ell }}^{\text{dec}}\backslash hat\{x\}^\{\backslash text\{out\}\}\_\backslash ell\; =\; \backslash sum\_\{k\; \backslash leq\; \backslash ell\}\; W^\{\backslash text\{dec\}\}\_\{k\; \backslash to\; \backslash ell\}\; f\_k\; +\; b^\{\backslash text\{dec\}\}\_\backslash ell$$

where $\sigma \backslash sigma$ is a sparsity-encouraging activation function.

The model is trained with a reconstruction loss alongside an auxiliary loss to encourage sparsity, roughly:

$$\mathcal{L}(x^{\text{in}},x^{\text{out}})=\mathrm{\parallel }{\widehat{x}}^{\text{out}}-x^{\text{out}}{\mathrm{\parallel }}_2^2+\lambda \mathrm{\parallel }\tanh (\alpha f){\mathrm{\parallel }}_1\backslash mathcal\; L(x^\{\backslash text\{in\}\},\; x^\{\backslash text\{out\}\})\; =\; \backslash |\backslash hat\{x\}^\{\backslash text\{out\}\}\; -\; x^\{\backslash text\{out\}\}\backslash |\_2^2\; +\; \backslash lambda\; \backslash |\backslash tanh(\backslash alpha\; f)\backslash |\_1$$

See our technical report for more details.

Model Details

This model is a cross-layer transcoder with 20480 features per layer (an expansion factor of 20x), using a JumpReLU activation function. It attains an L0 sparsity across layers of approximately 115, with about 23% of variance in MLP outputs unexplained.

The model was trained on approximately 750 million tokens of text from a broad range of domains, including general web text, public domain books, scientific articles, and code. Source datasets include:

• DCLM
• The Stack
• Common Corpus
• MathPile

Usage

This CLT can be used with the circuit-tracer library to generate attribution graphs.

Note: To load the CLT in circuit-tracer with the Qwen3-0.6B-Base model, you will need to install a patched version of Transformer Lens, as the current release does not support the Qwen3-Base models. Such a patched version is available here. The CLT can be used with an unpatched Transformer Lens if you use the Qwen/Qwen3-0.6B model instead.

You can load the CLT in circuit-tracer as follows:

import torch
from circuit_tracer import ReplacementModel
model_name = "Qwen/Qwen3-0.6B-Base" # Or just "Qwen/Qwen3-0.6B"
transcoder_name = "bluelightai/clt-qwen3-0.6b-base-20k"
clt = ReplacementModel.from_pretrained(model_name, transcoder_name, dtype=torch.bfloat16)

See this Colab notebook for a complete example.

Weight format

The model weights are sharded by layer across multiple files. The W_enc_{layer}.safetensors file contains the following named tensors:

• W_enc_{layer}: encoder weights, shape (d_latent, d_model)
• b_enc_{layer}: encoder biases, shape d_latent,
• threshold_{layer}: JumpReLU thresholds, shape d_latent
• b_dec_{layer}: decoder biases, shape d_model.

The W_dec_{layer}.safetensors file has a single tensor named W_dec_{layer}, containing the weights decoding that layer's features to the outputs of all subsequent layers, with shape (d_latent, n_out_layers, d_model).

Contact

You can reach us at [email protected] with any questions or inspiration.