Drowse

Drowse is a local workbench for mechanistic interpretability on large language models.

It comes with a local dashboard, along with a Python API, and a server compatible with both OpenAI and Ollama.

Quick start

The PyPI distribution name is drowse.ai; the Python import and CLI are drowse.

pip install drowse.ai
drowse serve google/gemma-3-4b-it

Open http://localhost:8000.

The first launch downloads the model and fits the 17 bundled concept probes. This can take a while; they get stored in ~/.drowse/ for future launches.

For NVIDIA CUDA:

pip install "drowse.ai[cuda,flash]"
drowse serve google/gemma-3-4b-it --device cuda

Browser edition

drowse.ai runs the same dashboard in the browser, with inference, steering, probes, and the R lens on the device through WebGPU and no server. Its source, the in-browser runtime, and its release tooling live in stdlogits/drowse.ai, which pins this repository as a submodule for the shared dashboard and the Python reference implementation. The Python server and the bundled dashboard described here are independent of it.

The dashboard

The header holds six rooms: Workbench, Units, Lens, Probes, Artifacts and Model, and its search (⌘K) finds any room, tool or kind of steering by what it does ("fit a lens", "steer a neuron"). The workbench is where you talk to the model: the thread (the path you are on through the conversation, and your earlier chats), the page (the conversation, or the loom: every turn as a node in a branching tree you can reroll, compare and fork from) and the Inspector. They sit side by side on a wide screen and fold into one column on a phone, with the Inspector as a sheet.

The Inspector shows what you selected. With nothing selected it holds the instruments, neurons first: the unit probes, the live lens readout, the geometry probes, the SAE's features, sampling, and one line for each source the lens and SAE read through.

• A token. Click any token for its alternatives with logprobs (and a fork from any of them), its anatomy (the neurons and heads that pushed it, each pushed or silenced in a click, then what the residual reads as through both lenses layer by layer and what attention and the MLP, and the steering, added) and its probe and feature readings. The readings can be replayed without the steering for comparison.
• A probe opens its fitted geometry, layer by layer.
• A unit (a neuron or head) opens what it writes and fires on, its label and its activity over the reply.

Steering is built from chips, in one place over the composer. + Steer adds a neuron or head (by ID or by a word it writes), a lens token, a concept, a manifold position or an SAE feature; the instruments' Push and Silence add one in a click, and the chip opens under the strip without leaving what you were looking at. A chip's editor sets a signed strength, the layers it acts at (a depth strip), when it acts (the prompt, the response, the first or all-but-the-first N tokens) and "Only while…", a gate on an attached probe's reading; a concept can also be kept along, or removed from, another (~, |), and any direction or unit can be removed (!). The dashboard never takes typed recipe text: a recipe it cannot draw as chips (from an agent or the API) shows as one read-only chip and is sent as written. Agents, the HTTP APIs, the CLI and Python keep the full expression language below.

A reply cut short by Stop (or a dropped connection) is kept and marked "stopped early"; a browser that reads slower than the model writes never cuts one, and catches up from the server's conversation tree. Saved chats live in the browser (saved automatically between replies) and as backup files you download; the server keeps no conversations.

The rooms beside the workbench read the conversation token by token. A strip of its tokens, prompt tokens included, sits at the top of each: step with ← and →, jump with Home and End, click a token, or read the whole reply. The workbench and the rooms share that cursor, so a token chosen in one is the token in the other.

Room	Purpose
Units	The neurons and heads strongest at the token (or at their peak over the reply), a whole layer at once, and the chosen unit's activity, writes, Push and Silence
Lens	Each layer at the token through both lenses (what the S lens reads, what the R lens says, weighted by depth), steered or not, with the lens probes
Probes	The linear probes and vectors (the value at the token, the trace over the conversation), then the manifolds (where the token sits in the fitted geometry), then SAE features
Artifacts	Where everything is prepared: the neuron-lens unit catalog, the lenses (fetched or fitted, both families at both sites), probes (built, installed, combined), templates, roles and SAEs
Model	The model's layers, neurons and heads, device and versions, what is prepared for it, health and API access

Chat mode renders the model's chat template as turns, with collapsible thinking, any role label a compatible template supports, and a cast that assigns steering to roles; raw mode is one buffer for base models.

Concepts, subspaces, and manifolds

In Drowse, you extract concepts as manifolds or subspaces.

• A 1D flat subspace is just a steering vector.
• A higher-rank flat subspace is a group of orthogonal steering vectors.
• A curved manifold fits a nonlinear surface and lets you steer along it.

Bundled concepts

Drowse comes with 17 concepts that are attached as probes by default: 14 two-pole axes and 3 one-pole concepts (deceptive, refusing, manipulative):

Category	Concepts
Epistemic	confident.uncertain, deceptive, curious.disinterested
Alignment	refusing, sycophantic.blunt, manipulative
Register	formal.casual, direct.indirect, verbose.concise, creative.conventional, humorous.serious, warm.clinical, technical.accessible
Cultural	masculine.feminine, individualist.collectivist, traditional.progressive, religious.secular

A one-pole concept is measured from the model's own default answers, which are already honest, compliant and sincere, so it has no opposite pole to write: a positive strength pushes toward it and a negative one past the default (-0.5 deceptive is more honest than the model's default).

Three larger concept sets ship as well, but aren't fitted by default:

• personas — 107 personas;
• emotions — 20 emotional states;
• months — 12 months.

Steering expressions

Every surface uses the same expression format. The same recipe can be used across Python, YAML, OpenAI, Ollama, or the native API.

-0.3 deceptive + 0.4 warm
0.5 formal - 0.2 verbose
0.3 deceptive|sycophantic.blunt
0.3 deceptive~confident.uncertain
!sycophantic.blunt
0.5 personas%pirate
0.7,0.4 months%january@after
0.4 warm@when:confident.uncertain>0.4
0.3 jlens/orange + 0.2 sae/9143
0.5 deceptive{12-20} + 0.4 personas%hacker{workspace}
!jlens/Paris{24-30}@response

Syntax	Meaning
+, -	Add or subtract terms
leading number	Steering coefficient; omitted terms default to 0.5
~	Keep the component shared with another direction (Mahalanobis)
|	Remove the component shared with another direction (Mahalanobis)
!	Mean-ablate a direction or a unit; a bare !x ablates fully
%label or %x,y,…	Choose a named node or coordinates on a manifold
{12-20}, {4-8,20}, {workspace}	Act only at these layers (after the selector, before @)
@prompt, @response, @after, @thinking, …	Restrict the token phase where a term applies
@first:N, @after:N	Restrict a term to a counted decode window
@when:<probe><op><value>	Apply a term only while a live probe gate is true
jlens/<word>, jlens/#<id>, sae/<id>	Steer toward saying a token (its R-lens direction) or along an SAE decoder row
slens/<word>, slens/#<id>	Steer toward having read a token: its S-lens footprint at each layer
neuron/L8n344, head/L20h17	A unit: push along a neuron's write, or mean-ablate a neuron or head (!head/L20h17)

A bare pole or node label such as formal or pirate resolves to its concept or manifold; an ambiguous name raises instead of guessing. The operands of ~, | and ! must reduce to one direction: a whole two-pole concept (sycophantic.blunt, not sycophantic), a one-pole concept (deceptive), an extracted vector, or a jlens//slens//sae/ atom (! also takes a unit). A head has no push direction, so head/ takes only !. @response covers every generated token, thinking included; @after applies only after the thinking block closes. A gate names an attached probe, so confident.uncertain above reads the bundled probe. The larger sets (personas, emotions, months) steer once they are fitted for the model.

Without a layer clause a term acts on every layer it is fitted at. A clause names layers ({12-20}, {8,10,12}), the model's 40–90% depth band ({workspace}, intersected with the fitted layers) or {all}; naming a layer the term is not fitted at is an error. The coefficient keeps its meaning under a clause: a one-layer selection steers about as hard per layer as the whole fit. neuron/ and head/ atoms carry their own layer and take no clause.

Manifold coefficients use two coordinates: along and onto. along controls movement within the manifold toward the target; onto reduces the off-surface component inside the manifold's fitted tube. For a flat subspace, aim for a coefficient near 0.5 and raise it toward 1.0 for stronger expression.

How Drowse works

Extraction

Drowse first has the model answer a shared set of baseline prompts as each concept, then it takes the resulting hidden states and fits them to either a curved manifold or a flat subspace.

Layer allocation uses a Mahalanobis metric estimated from neutral activations. Discriminative layer selection removes flat axes that fail to straddle the neutral baseline across the fitted nodes.

The full data flow, artifact boundaries, instrument protocol, and concurrency invariants are documented in ARCHITECTURE.md.

Monitoring

A reading includes fitted coordinates, the centered activation's subspace fraction, the nearest nodes and their soft assignment, and, for curved manifolds, an off-surface residual and tube membership.

Jacobian lenses

The Jacobian lens implementation follows Gurnee et al.'s work. Drowse lets you use one of the published lens artifacts, or fit your own.

A lens reads a vector inside the model as a ranking over tokens, in one of two ways. The R-lens (or the standard J-lens) reads what the vector is disposed to make the model say; the S-lens (or the standard K-lens) reads which input tokens it looks like it has read. Each comes at two places in a layer: out, the block's output (also where an MLP neuron writes and where a head reads, one layer up), and mid, the residual after attention (where a neuron reads and a head writes). Which one to trust is a question of depth: the S read is the faithful one early in the model, the R read late, and both show, weighted, in between. drowse lens top reads every layer of a prompt that way and labels each read with the lens it used; session.lens.read(vector, layer, site, of=...) does the same for any vector. The S read reports a footprint z, each token's score as a z-score over the vocabulary. A residual (the stream itself) carries a large component every position shares, so its S read subtracts the model's neutral layer mean first (of="residual"); a direction such as a steering vector or a neuron's write is read as it is (of="direction").

stdlogits/rs-lenses publishes both lenses at both places for Qwen3-8B and Gemma 4 12B (lens fetch MODEL rs-r and rs-s). Drowse keeps one R-family and one S-family lens selected per model. jlens/<word> steers toward saying a token and slens/<word> toward having read it; as probes, jlens/ reads the mean readout probability and slens/ the footprint z averaged over the layers where the S lens is the faithful read (weighted by depth, early layers most), and either can gate a term (@when:slens/owl > 3).

An S read ranks only tokens the model saw often enough in text: a token it almost never saw has an untrained embedding that would top early reads. The counts come from the neuron-lens catalog when it is bound (unit fetch), else from drowse lens counts MODEL, which streams the S lens's corpus through the tokenizer (no model load); without either, the tokenizer's own rules decide, and lens show and lens top say so.

Neurons and heads

Every MLP neuron (neuron/L8n344) and attention head (head/L20h17) is an object Drowse reads, probes, gates on and steers. drowse unit read says what one is about from its weights alone, through the same depth rule as the lens reads: a neuron per pole (its activation takes both signs) writes (its output direction, read where it lands, at the block output) and fires on (its input direction, read at the MLP input); a head copies or suppresses tokens (its OV circuit) and has a query, key and value that pick up tokens at the layer below. Hub tokens that top many reads at once from the lens's geometry can be hidden (--hide-hubs), which ranks each token against the rest of the layer. drowse unit top owl --by writes finds the neurons that write a token most specifically, and drowse unit find deceptive the neurons a steering direction is made of.

For Qwen3-8B and Gemma 4 12B the neuron-lens catalog adds a label to every neuron pole and head, written by Claude Haiku 5.5, and how often each fires; drowse unit fetch MODEL binds it (labels and small tables only, pinned to a commit), and every unit view then links to the unit's page on neuron-lens.com. A unit's readings then divide by the catalog's typical high firing, so 1.0 means "as strong as this unit usually fires in text"; without a catalog they divide by its typical firing on Drowse's neutral statements, a smaller sample, so the same firing reads several times larger, and every view says which.

Sparse autoencoders

Drowse can either use a published SAELens release or train a local SAE.

Installation

Drowse requires Python 3.11 or newer and PyTorch 2.13 or newer. CUDA or Apple Silicon MPS is strongly recommended for interactive use; CPU is supported for smaller models and non-GPU workflows.

pip install drowse.ai

The base package includes the HTTP server and the prebuilt Svelte WebUI. Optional extras add specialized workflows:

Extra	Adds
flash	FlashAttention 2 for supported NVIDIA CUDA models; stable and tested
cuda	bitsandbytes quantization and Hugging Face kernels acceleration
sae	SAELens 6.x, for published SAE releases (sae fetch); local SAEs need no extra
hf	datasets for streamed lens and SAE corpora
gguf	GGUF import/export support
research	datasets, NumPy, SciPy, scikit-learn, pandas, Matplotlib, and image helpers
notebook	Plotly, pandas, and Kaleido notebook helpers
dev	Test, lint, type-check, and build tooling

Extras can be combined:

pip install "drowse.ai[cuda,flash]"     # full tested NVIDIA path
pip install "drowse.ai[sae,hf]"         # SAELens sources and streamed corpora
pip install "drowse.ai[hf,research]"    # dataset-backed research workflows
pip install "drowse.ai[notebook]"       # interactive figures

cuda and flash are Linux/NVIDIA CUDA extras. FlashAttention is selected automatically when installed; there is no runtime flag to enable it.

From source:

uv pip install -e ".[dev]"

Running the server

drowse serve MODEL [options]

Repository Python is disabled by default. For a trusted model that requires custom code, use DROWSE_TRUST_REMOTE_CODE=1 drowse serve MODEL; Python callers can pass trust_remote_code=True to DrowseSession.from_pretrained. This grants the model repository permission to execute Python locally.

Common options:

Option	Default	Purpose
-d, --device	auto	cuda, mps, cpu, or automatic selection
-q, --quantize	none	4bit or 8bit bitsandbytes quantization on CUDA
--dtype	device default	float32, bfloat16 or float16 weights (default bf16 on CUDA/MPS, fp32 on CPU); a fit records the dtype it ran at
-p, --probes	all	Bundled probe categories, all, or none
-H, --host	127.0.0.1	Bind address; non-loopback requires an API key
-P, --port	8000	Bind port
-S, --steer	none	Default steering expression
-C, --cors	none	Allowed CORS origin; repeatable
--top-k-alts	0	Alternative tokens captured at each decode step
--compile	off	Opt into torch.compile after Drowse probes the path
--cuda-graphs	off	Pair static cache and CUDA graph capture with --compile
-k, --api-key	none	Require bearer authentication; also reads $DROWSE_API_KEY
--no-web	off	Run the APIs without mounting the dashboard
-c, --config	none	Setup YAML; repeatable, later files win

serve and every subcommand that accepts -c/--config read ~/.drowse/config.yaml first and then compose any explicit -c PATH files on top. For example:

model: google/gemma-3-4b-it
vectors: "-0.3 deceptive + 0.2 warm"
temperature: 0.8
top_p: 0.9
max_tokens: 512
return_top_k: 8

Command-line flags override YAML, and vectors: sets the server's default steering only; the experiment verbs use exactly their own -S. Inspect the resolved configuration with drowse config show and validate a file with drowse config validate path.yaml.

Command-line artifact workflows

The CLI has nine top-level verbs. The WebUI covers the interactive versions of most workflows; the CLI is useful for reproducible preparation, distribution, and batch work.

Verb	Role
serve	Launch the WebUI and the three HTTP protocol surfaces
manifold	Extract, generate, derive from a template, fit, bake, merge, transfer, compare, or diagnose manifolds
pack	List, inspect, install, search, push, clear, refresh, remove, or export manifold packs
experiment	Run alpha fans, replay transcripts, and evaluate naturalness
config	Show or validate composed configuration
template	Create and score restricted-choice completion templates
lens	Fit, fetch, select, read, decompose, or remove Jacobian lenses
sae	Train, fetch, select, inspect, or remove SAE sources
unit	Read neurons and heads from their weights, find the neurons behind a token or a direction, and fetch, list, or remove the neuron-lens catalog

Representative commands:

# Extract and fit a two-pole concept for one model
drowse manifold extract patient impatient -m google/gemma-3-4b-it

# Fit a bundled many-node manifold
drowse manifold fit personas -m google/gemma-3-4b-it

# Install or publish manifold packs through Hugging Face
drowse pack search creativity
drowse pack install OWNER/REPO
drowse pack push local/patient.impatient -a OWNER/REPO -m google/gemma-3-4b-it

# Move one manifold closure, plus its referenced template, as an archive
drowse pack export archive local/patient.impatient -o patient.drowse
drowse pack install patient.drowse

# Fetch a provider lens (Neuronpedia's J-lens by default), or fit a local one:
# --kind r (the default) is the RelP R-lens at local:r, --kind j the standard
# J-lens; --kind s and k are the S family's S- and K-lenses. Each fits the
# block output (out) and the MLP input (mid) unless --sites says otherwise
drowse lens fetch google/gemma-3-4b-it
drowse lens fit org/model
drowse lens fit org/model --kind s --prompts 100
drowse lens fit org/model --kind j --sites out --layers workspace
# Or on another commit-pinned Hub dataset (a shuffled one grows by nested samples)
drowse lens fit org/model --hf-corpus HuggingFaceFW/dclm_100BT-shuffled --prompts 1000
# A bilingual mix: 3 English documents per Chinese one, the Chinese side spread
# over 64 of its (unshuffled) data files
drowse lens fit org/model --prompts 1000 \
  --hf-corpus HuggingFaceFW/dclm_100BT-shuffled,weight=3 \
  --hf-corpus HuggingFaceFW/fineweb-2:cmn_Hani,files=64
# Pick which lens the session reads (one R-family and one S-family source)
drowse lens use org/model local:r
# rs-lenses' R- and S-lens, both sites, for Qwen3-8B and Gemma 4 12B
drowse lens fetch Qwen/Qwen3-8B rs-r
drowse lens fetch Qwen/Qwen3-8B rs-s
# Without a neuron-lens catalog, count the S lens's corpus for the S token mask
drowse lens counts org/model
# Read a prompt's residuals through both (--site mid: after each attention)
drowse lens top Qwen/Qwen3-8B "The capital of France is" --site out
# Which input tokens' footprints a direction resembles
drowse lens decompose deceptive -m Qwen/Qwen3-8B --dictionary tokens-in

# Neurons and heads: what one is about, who writes a token, what a direction is made of
drowse unit fetch Qwen/Qwen3-8B                      # the neuron-lens labels (Qwen3-8B, Gemma 4 12B)
drowse unit read neuron/L8n344 -m Qwen/Qwen3-8B --hide-hubs
drowse unit read head/L20h17 -m Qwen/Qwen3-8B
drowse unit top owl -m Qwen/Qwen3-8B --by fires --layers 0-12
drowse unit find deceptive -m Qwen/Qwen3-8B

# Fetch an SAE (needs the [sae] extra) or train a local source
drowse sae fetch google/gemma-3-4b-it saelens:gemma-scope-2-4b-it-res
drowse sae train org/model my-sae --layer 20 --tokens 1000000

A local fit averages 64 documents by default and prints its split-half difference per site (how differently two halves of the corpus read: the signal for whether more documents would help). Fits run on dense models (llama, mistral, qwen2/3, gemma 2/3/4 text); mixture-of-experts and hybrid stacks are refused, though their published lenses still load. Lens fits are resumable: a fit checkpoints every 25 documents and every ten minutes, Ctrl-C or SIGTERM finishes the document in flight and saves before exiting (a second signal stops at once), rerunning the command continues, and an exact repeat is recognized without loading the model. A fit never silently replaces a local lens over more documents or more layers: a 5.4 lens fitted with the old default of 100 documents needs --prompts 100 (which adds the mid site) or -f to replace it. Run drowse <verb> -h and drowse <verb> <subcommand> -h for the complete flag surface.

HTTP APIs

The same drowse serve process exposes four surfaces on one port:

• / — the Drowse WebUI;
• /v1/* — OpenAI-compatible models and chat completions;
• /api/* — Ollama-compatible generation and chat;
• /drowse/v1/* — native sessions, loom trees, probes, manifolds, templates, lens, SAE and unit lifecycles and reads, replay, SSE, and token-plus-measurement WebSockets.

Interactive OpenAPI documentation is available at http://localhost:8000/docs.

OpenAI client

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="unused")
response = client.chat.completions.create(
    model="google/gemma-3-4b-it",
    messages=[{"role": "user", "content": "Describe a rainy afternoon."}],
    extra_body={"steering": "0.3 warm + 0.2 concise"},
)
print(response.choices[0].message.content)

Ollama client

curl -N http://localhost:8000/api/chat -d '{
  "model": "gemma3",
  "messages": [{"role": "user", "content": "Write a short haiku."}],
  "options": {"steer": "0.3 warm - 0.2 formal.casual"}
}'

Drowse targets a trusted local machine or lab network. It is not a hardened multi-tenant inference service. If you bind it beyond a trusted host, read SECURITY.md, set an API key, and add TLS, rate limits, request limits, and isolation outside Drowse.

Python API

from drowse import DrowseSession, SamplingConfig

with DrowseSession.from_pretrained(
    "google/gemma-3-4b-it",
    device="auto",
    return_top_k=8,
) as session:
    name, profile = session.extract("patient", baseline="impatient")
    session.add_probe(name)

    result = session.generate(
        "How should I learn a difficult skill?",
        steering=f"0.3 {name} + 0.2 concise",
        sampling=SamplingConfig(
            temperature=0.8,
            top_p=0.9,
            max_tokens=256,
            seed=42,
        ),
    ).first

    print(result.text)
    print(result.applied_steering)
    print(result.probe_readings[name].coords)

generate and generate_stream accept the same steering expression as the WebUI; session.steering(expr) applies one to every generation inside a with block. Generation returns a list-like RunSet; .first is convenient for a single completion. GenerationResult carries text, token IDs, throughput and timing, finish reason, the canonical applied expression, captured log probabilities, aggregate probe readings, and the versioned measurement envelope every instrument family writes into.

Batch and sweep helpers return the same result shape:

batch = session.generate_batch(
    ["Describe a sunset.", "Describe a storm."],
    steering="0.3 warm",
)

sweep = session.generate_sweep(
    "Describe a forest.",
    sweep={"warm.clinical": [-0.4, 0.0, 0.4]},
)

Restricted-choice scoring reads the model's raw probability distribution over a set of candidate completions, optionally under steering, so you can see whether steering moved the distribution and not just the top choice:

scores = session.score_choices(
    [{"role": "user", "content": "The first weekday is"}],
    ["Monday", "Tuesday", "Wednesday"],
    steering="0.3 confident",
)

The four read families are available as session.geometry, session.lens, session.sae and session.units; session.add_probe attaches a concept, a jlens/<word> or slens/<word> token probe, an sae/<id> feature or a neuron/L8n344 / head/L20h17 unit to the right one. session.lens.read(vector, layer, site, of="residual" | "direction") reads any vector through both lens families at once.

session.token_anatomy(node_id, raw_index) replays one token and reads it layer by layer: what the residual reads as through both lenses, what attention and the MLP (and any steering) added, the neurons and heads that pushed the token through the R lens, and where the top heads looked. session.unit_activity(node_id) reads every neuron and head over a turn's tokens: the strongest units with their peak token, per-token traces of named units, and one token's whole population. Both read a prompt turn too (its tokens replay through the turn's recorded capture). Both are also served at /drowse/v1/sessions/{id}/instruments/token-anatomy and .../instruments/units/activity.

Notebook helpers are available from drowse.notebook after installing drowse.ai[notebook]: plot_alpha_sweep, plot_probe_correlation, plot_layer_norms, plot_trait_history, and to_dataframe.

Model support

Drowse has end-to-end tested paths for:

• Qwen 2, Qwen 3, and Qwen 3.5, including its text and MoE variants;
• Gemma 2, Gemma 3, and Gemma 4, including text-only extraction from supported multimodal checkpoints;
• Mistral 3 and Ministral 3;
• Llama, GLM, gpt-oss, and Talkie.

Other decoder-only models load when their transformer blocks sit at a standard Transformers location, which covers most families (Mixtral, Phi, Cohere, DeepSeek, OLMo, Granite, GPT-2, Falcon, MPT, OPT, and others). Drowse warns once when an architecture has not been exercised end to end.

CUDA and Apple Silicon MPS both have real-model smoke coverage. Model-specific features still depend on the checkpoint: role-augmented extraction and persona regeneration need a chat template with a substitutable assistant label (Mistral's does not have one), published SAE and lens sources cover only some models, lens fits and unit reads refuse mixture-of-experts and hybrid stacks, and FlashAttention depends on the model's Transformers attention implementation.

State and distribution

Drowse keeps local state under ~/.drowse/; set $DROWSE_HOME to move it. The store contains authored manifolds, per-model fits and neutral statistics, local lens and SAE artifacts, source and catalog bindings, and templates. It holds no conversations: the dashboard keeps its saved chats in the browser and as downloaded backup files, and Python saves a tree only to a LoomTree.save() path you choose.

Manifold packs are folders with metadata, node corpora, integrity hashes, and optional fitted tensors. They can be installed from a local path or distributed as Hugging Face model repositories. A .drowse is the narrow ZIP transport for exactly one manifold closure and its referenced template, when present; export it with drowse pack export archive and install it with drowse pack install. It never contains model weights, lenses, or SAEs. A fitted two-node PCA manifold can instead be exported as a llama.cpp control-vector GGUF with drowse pack export gguf.

A Hugging Face pack is a model repository tagged drowse-manifold that holds exactly one .drowse at its root, a root manifold.json summary, and a model card; drowse pack push writes that layout, and both drowse and the drowse.ai browser edition install from it. An install pins the repository to an immutable commit and records that commit as the manifold's source; the browser also checks that the archive's declared repository and revision match it.

Treat model repositories and downloaded artifacts as executable or otherwise untrusted input. Drowse validates archive paths, declared structure, tensor headers, and hashes, but integrity is not publisher identity.

Drowse reads only its current artifact formats and does not migrate older ones. Releases through 5.3.0 were published as saklas, and drowse does not read a ~/.saklas directory. Lens artifacts fitted or fetched before 5.4 must be converted once from a source checkout with python scripts/migrate_lens_v7.py --home ~/.drowse (try --dry-run first). 6.0 recomputes its neutral statistics on first load and does not read 5.x saved conversations or transcripts; manifolds, templates, packs and lenses carry over unchanged.

Development

pip install -e ".[dev]"

ruff check .
pyright
pytest -q -m "not gpu"              # the CI suite
pytest -q -n auto -m "not perf"     # everything this machine can run
pytest -m perf -p no:xdist          # throughput gates, serially
python -m build

GPU integration tests download the public SmolLM2-360M-Instruct weights by default. Set DROWSE_TEST_MODEL to exercise another compatible model:

pytest -q tests/gpu/test_smoke.py
DROWSE_TEST_MODEL=google/gemma-3-4b-it pytest -q tests/gpu/test_smoke.py

The WebUI is a Svelte 5 + Vite application in webui/. Its compiled bundle under drowse/web/dist/ is committed package data and ships in the wheel.

cd webui
npm ci
npm run check
npm run build
git diff --exit-code ../drowse/web/dist

See CONTRIBUTING.md for development conventions and adding a model architecture.

Research lineage and credits

Drowse builds on Representation Engineering (Zou et al., 2023). repeng by Theia Vogel is the best-known compact implementation of that approach; Drowse takes the workbench route, adding live monitoring, manifold geometry, branching experiments, and server protocols.

Two-pole extraction uses difference-of-means following Im & Li, 2025. Manifold steering follows Goodfire's manifold work. The personas source is derived from the framing in Anthropic's Assistant Axis paper. The Jacobian lens implements the verbalizable-workspace method of Gurnee et al., 2026. Unit reads follow neuron-lens's definitions (its layer-contrastive reads, Elhage et al.'s copying score, Gurnee et al. 2024's prediction neurons, Stolfo et al. 2024's entropy neurons); its catalog's labels are by Claude Haiku 5.5.

Drowse is written by a9lim with Claude. The dashboard is drawn in Ledger, the Standard Logits design system.

If you use Drowse in published research, please cite the relevant upstream methods alongside the Drowse version and exact model checkpoint you used.

Contact, issues, and security

For questions, feedback, or research inquiries, email contact@drowse.ai or use the contact form.

Please update to the latest Drowse release before filing a bug. Include the model ID, device, dtype or quantization mode, Drowse version, and a minimal reproduction in GitHub Issues.

Report vulnerabilities privately according to SECURITY.md.

License

Drowse is licensed under AGPL-3.0-or-later. See LICENSE.