Skip to content
Other

statetrace-qwen3.5-0.8b-transcoders

per-layer transcoders for dissecting Qwen3.5-0.8B

2 min read

A set of 24 per-layer transcoders (PLTs — modules for interpreting a model's internal computations) by Fealtyy, built for analyzing the Qwen/Qwen3.5-0.8B language model. The base model uses a hybrid architecture with 18 Gated DeltaNet layers and 6 classic full-attention layers. Each transcoder reads an MLP block's input and predicts its output across 16,384 sparse JumpReLU features, letting you build attribution graphs (circuit tracing) and track decision logic inside the network. This tool is aimed at mechanistic interpretability researchers working with the statetrace and circuit-tracer libraries.

Key features

  • Covers all 24 layers of Qwen3.5-0.8B with a width of 16,384 features per layer (16x the 1024 residual stream dimension).
  • Sparsity of ~48 active features per token (L0) and 0% dead features thanks to the AuxK loss function.
  • Fully compatible with the circuit-tracer transcoder format: weight matrices are saved in layer_{i}.safetensors at fp32 precision alongside config.yaml.
  • High attribution graph accuracy on test prompts, averaging a 0.72 replacement score and a 0.93 completeness score.

What you can do

  • Build attribution graphs for short prompts in statetrace to see which internal features fire during response generation.
  • Compare how the hybrid Gated DeltaNet layers behave versus standard full-attention layers (layers 3, 7, 11, 15, 19, and 23) inside Qwen3.5-0.8B.
  • Load individual layers directly via safetensors for custom MLP feature labeling and analysis experiments.

Example

from statetrace.replacement import ReplacementModel
from statetrace.attribution import attribute

rm = ReplacementModel.from_pretrained(
    "Qwen/Qwen3.5-0.8B",
    "Fealtyy/statetrace-qwen3.5-0.8b-transcoders",
    local_files_only=False,
)
graph = attribute("The Eiffel Tower is located in the city of", rm)
Read moreDetails
Trained on 399.6M tokens per layer from the FineWeb-Edu dataset (1024-token context) in 3.5 hours on a single H200 GPU (~25 GB).
Metrics on 32,736 validation tokens: baseline model cross-entropy is 2.855, rising to 3.464 (+0.61 nats) when swapping in all 24 transcoders simultaneously; mean FVU across layers is 0.230 (ranging from 0.070 on layer 23 to 0.355 on layer 9).
Limitations: trained exclusively on English educational web text (performance on code, chat, and other languages was not measured); features do not include automated text labels; these are per-layer (PLT) rather than cross-layer (CLT) transcoders.
Discuss Telegram 0

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # Fealtyy/statetrace-qwen3.5-0.8b-transcoders
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download Fealtyy/statetrace-qwen3.5-0.8b-transcoders --local-dir ~/models/statetrace-qwen3.5-0.8b-transcoders
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of statetrace-qwen3.5-0.8b-transcoders (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

Fealtyy/statetrace-qwen3.5-0.8b-transcoders
Author
Fealtyy
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments