Qwen3.5-4B PLTs

Per-layer PLT checkpoints for Qwen/Qwen3.5-4B trained as same-layer MLP transcoders.

Contents

  • transcoder_L0.pt
  • transcoder_L2.pt ... transcoder_L31.pt
  • layer_metrics.json

Layer L1 is currently missing because that checkpoint was not preserved in the training artifacts.

Training recipe

  • Base model: Qwen/Qwen3.5-4B
  • Objective: same-layer MLP input to same-layer MLP output
  • Feature width: 16384
  • Activation regime: ReLU + L1 with ghost gradients and dead-feature resampling
  • Training corpus: 20k image-text subset built from CC3M and CC12M

Intended use

These checkpoints are intended for research on sparse feature discovery, circuit analysis, and PLT-based steering experiments in Qwen3.5-4B.

Notes

Checkpoint quality is not uniform across layers. Later layers generally behaved better than early layers in our experiments.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for KokosDev/qwen35-4b-plt

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(87)
this model