recurrentlens 0.1.0.post1__tar.gz

recurrentlens-0.1.0.post1/.github/workflows/ci.yml

@@ -0,0 +1,51 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+      - name: Install uv
+        run: pip install uv
+      - name: Install ruff
+        run: uv pip install --system ruff
+      - name: Ruff format check
+        run: ruff format --check src tests
+      - name: Ruff lint
+        run: ruff check src tests
+
+  no-secrets:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Grep for token-shaped strings
+        run: bash scripts/check_no_secrets.sh
+
+  unit-cpu:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python: ['3.11', '3.12']
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python }}
+      - name: Install uv
+        run: pip install uv
+      - name: Install package + dev deps (CPU)
+        run: |
+          uv pip install --system -e ".[dev]"
+      - name: Pytest (smoke + unit, CPU)
+        env:
+          RECURRENTLENS_SKIP_GPU_TESTS: '1'
+        run: pytest -ra

recurrentlens-0.1.0.post1/.gitignore

@@ -0,0 +1,66 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# venv
+.venv/
+venv/
+ENV/
+env/
+.env
+
+# Testing / coverage
+.coverage
+.coverage.*
+.pytest_cache/
+htmlcov/
+.tox/
+.nox/
+*.cover
+coverage.xml
+
+# IDE
+.vscode/
+.idea/
+*.swp
+*.swo
+.DS_Store
+
+# Type checkers / linters
+.mypy_cache/
+.pyright/
+.ruff_cache/
+
+# Notebooks
+.ipynb_checkpoints/
+
+# uv
+uv.lock
+
+# Project-specific
+artifacts/
+activations_cache/
+experiments/_wip/
+*.zarr/
+*.safetensors
+*.pt
+*.html

recurrentlens-0.1.0.post1/.pre-commit-config.yaml

@@ -0,0 +1,14 @@
+repos:
+  - repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.6.9
+    hooks:
+      - id: ruff
+        args: [--fix]
+      - id: ruff-format
+  - repo: local
+    hooks:
+      - id: no-secrets
+        name: Grep for token-shaped strings
+        entry: bash scripts/check_no_secrets.sh
+        language: system
+        pass_filenames: false

recurrentlens-0.1.0.post1/CHANGELOG.md

@@ -0,0 +1,88 @@
+# Changelog
+
+All notable changes follow [Keep a Changelog](https://keepachangelog.com/en/1.1.0/)
+and the project version follows [PEP 440](https://peps.python.org/pep-0440/)
+(SemVer-compatible, with `post` releases reserved for correctness-only patches).
+
+## [0.1.0.post1] — 2026-05-23
+
+Correctness hotfix on top of v0.1.0. No new features, no API additions, no
+breaking changes to the eight public functions. The release fixes silent
+correctness bugs uncovered in a post-release multi-agent audit.
+
+### Fixed
+- **JumpReLU SAE — straight-through estimator** (`sae/variants.py`).
+  The forward gate is now `z = z_pre * gate`, the canonical STE form. The
+  v0.1.0 implementation multiplied the surrogate gradient by `0.0`, which
+  silently severed the backward path. `JumpReLUSAE` is now trainable.
+- **`bench.evaluate` — removed `ce_recovery_proxy`** (`bench/evaluate.py`).
+  The v0.1.0 ratio `ce_clean / ce_sae` was not the standard CE-recovery
+  formula and produced numbers that could not be compared against SAELens,
+  Anthropic, or DeepMind reports. The metric has been replaced by
+  `ce_delta = ce_sae − ce_clean` (lower is better, 0 is perfect). The full
+  zero-ablation baseline ships in v0.1.1.
+- **`viz.feature_explorer` — HTML escape**.
+  All user-controlled fields (`context`, `model_id`, `hook_site`, `variant`)
+  are now passed through `html.escape`. This closes an XSS vector for
+  feature pages rendered from Hub-loaded SAEs.
+- **`BaseSAE.load` — metadata validation** (`sae/base.py`).
+  Refuse to construct an SAE when the safetensors metadata is inconsistent
+  with the stored tensors (e.g., a lying `d_sae` field) or names an unknown
+  variant. Closes a supply-chain footgun for `hub.load_sae`.
+- **L0 metric — signed nonzero count** (`sae/base.py`, `bench/evaluate.py`).
+  `(z != 0)` instead of `(z > 0)`. Matches the standard SAE convention so
+  negative activations (JumpReLU / Vanilla) are counted correctly.
+- **`ssm_h_t` hook — warn once per module** (`hooks/registry.py`).
+  The proxy warning previously fired on every `_resolve_target` call (i.e.
+  every forward pass for ablate/steer/eval). It now fires once per layer
+  module.
+- **`train._cosine_lr` off-by-one** (`sae/train.py`).
+  The final training step no longer collapses to `LR=0`; the schedule floors
+  at 1% of `base_lr` at progress=1.
+- **`features.ablate` — `torch.no_grad()`** (`features/ablate.py`).
+  SAE encode/decode inside the intervention hook no longer builds an
+  autograd graph.
+- **`ActivationCache(mmap)` — overwrite warning** (`sae/cache.py`).
+  Truncating an existing `.memmap` file now emits a `UserWarning` instead
+  of silently destroying data.
+- **`hub.push_sae` — `tempfile.TemporaryDirectory`** (`hub/io.py`).
+  No more stale `/tmp/<repo_id>/...` directories left behind after upload.
+
+### Changed
+- **`hooks.resolve_target`** is now a public symbol (`recurrentlens.hooks`).
+  The previous `_resolve_target` private name is kept as a back-compat alias
+  until v0.2.
+- **README `Quick look`**: `n_tokens=200_000` (CPU smoke). The 100M-token
+  research-grade target is documented as the v0.1.1 pretrained-artifact path.
+- **README SAELens characterization**: factual link to
+  [decoderesearch/SAELens#311](https://github.com/decoderesearch/SAELens/issues/311)
+  with the open-since date; removed editorial "broken" wording.
+- **README** now carries an `Alpha` status badge near the top.
+- **`notebooks/02_explore_pretrained.ipynb`**: `REPO_ID = None` by default,
+  falls back to an untrained `TopKSAE` for the API demo. Pretrained
+  Hub artifacts ship in v0.1.1.
+
+### Added
+- `tests/test_sae_unit.py::test_jumprelu_gradient_flows_to_encoder`
+- `tests/test_sae_unit.py::test_l0_counts_signed_nonzero`
+- `tests/test_sae_unit.py::test_base_sae_load_rejects_inconsistent_meta`
+- `tests/test_sae_unit.py::test_base_sae_load_rejects_unknown_variant`
+- `tests/test_viz_bench.py::test_feature_explorer_escapes_xss`
+- `tests/test_viz_bench.py::test_feature_explorer_no_acts_emits_warning_banner`
+- `tests/test_viz_bench.py::test_evaluate_does_not_emit_ce_recovery_proxy`
+- `tests/test_hooks_unit.py::test_ssm_h_t_warning_is_emitted_once_per_module`
+- `tests/test_train_smoke.py::test_cosine_lr_final_step_above_zero`
+- `tests/test_features.py::test_ablate_zero_removes_feature_contribution`
+  upgraded from a shape-only tautology to an analytic-delta assertion.
+
+### Deferred to v0.1.1
+- Pretrained Mamba-130M SAEs at layers {2, 6, 10} via Colab T4.
+- TopK auxiliary-K dead-feature loss (Gao et al. 2024 §3.2).
+- True CE recovery formula with zero-ablation baseline.
+- Tangent-projected decoder gradient + checkpoint save in `train_sae_full`.
+- nnsight backend for exact `h_t` capture.
+
+## [0.1.0] — 2026-05-23
+
+Initial public release. SSM-first SAE + feature-viz + Hub registry framework
+for Mamba and Mamba-2. Apache-2.0, 38 unit tests, 3 Colab notebooks.

recurrentlens-0.1.0.post1/CITATION.cff

@@ -0,0 +1,29 @@
+cff-version: 1.2.0
+message: "If you use recurrentlens in academic work, please cite it as below."
+title: "recurrentlens: mechanistic interpretability for State-Space Models"
+authors:
+  - family-names: "hinanohart"
+version: 0.1.0.post1
+date-released: 2026-05-23
+license: Apache-2.0
+repository-code: "https://github.com/hinanohart/recurrentlens"
+keywords:
+  - mechanistic-interpretability
+  - sparse-autoencoder
+  - state-space-models
+  - mamba
+notes: |
+  v0.1.0.post1 ships a framework + scaffold + smoke + 3 Colab notebooks.
+  Pretrained Mamba-130M SAE artifacts land in v0.1.1 (community Colab T4).
+
+  The shipped train_sae default is n_tokens=200_000, sized for a CPU smoke
+  run. The research-grade design target is 100M+ tokens; v0.1.1 pretrained
+  artifacts will be trained at that scale.
+
+  v0.1.0.post1 is a correctness hotfix on top of v0.1.0:
+    - JumpReLU straight-through estimator (was a `* 0.0` no-op in v0.1.0).
+    - bench.evaluate: removed the ce_recovery_proxy ratio (added ce_delta).
+    - viz.feature_explorer: HTML-escape user-controlled fields (XSS guard).
+    - hub.load_sae: validate safetensors metadata before construction.
+    - L0 metric: count (z != 0), matching the standard SAE convention.
+  See CHANGELOG.md for the full list.

recurrentlens-0.1.0.post1/CONTRIBUTING.md

@@ -0,0 +1,47 @@
+# Contributing
+
+Thanks for your interest in `recurrentlens`. The library is in early alpha (v0.1.0); contributions are very welcome.
+
+## Setup
+
+```bash
+git clone https://github.com/hinanohart/recurrentlens.git
+cd recurrentlens
+uv venv
+source .venv/bin/activate
+uv pip install -e ".[dev]"
+pre-commit install
+```
+
+## Running tests
+
+```bash
+pytest -ra                           # full suite
+pytest -ra -k smoke                  # smoke only
+RECURRENTLENS_SKIP_GPU_TESTS=1 pytest -ra   # CPU-only (CI default)
+```
+
+## Style
+
+- `ruff format`
+- `ruff check --fix`
+- `pyright` (standard mode)
+
+Pre-commit hooks enforce all three plus a `no-secrets` grep gate.
+
+## High-leverage contributions for v0.1.0 → v0.1.1
+
+1. **Pretrained SAE artifacts**: run `notebooks/03_train_mamba130m_sae.ipynb` on Colab T4, then `recurrentlens.hub.push_sae(...)` to `hinanohart/recurrentlens-mamba130m-L{2,6,10}-sae`. PR adding the Hub IDs to README.
+2. **Hook sites for other SSMs**: RWKV-7, Jamba block-local hooks are v0.2 but design notes in `docs/hooks_design.md` welcome.
+3. **Feature explorer richness**: v0.1.0 ships a minimal HTML table; richer interactivity (highlighting, k-way comparison) for v0.1.x is open.
+
+## v0.2 roadmap
+
+- `extract_circuit` (recurrent-state-aware patching)
+- `diff.compare(sae_mamba, sae_gpt2)`
+- Falcon-Mamba-7B Hub registry
+- RWKV-7 / Jamba support
+
+## Code of conduct
+
+Be kind. Disagree about technical claims with citations.

recurrentlens-0.1.0.post1/LICENSE

@@ -0,0 +1,216 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+
+   Copyright 2026 hinanohart
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

recurrentlens-0.1.0.post1/PKG-INFO

@@ -0,0 +1,148 @@
+Metadata-Version: 2.4
+Name: recurrentlens
+Version: 0.1.0.post1
+Summary: Mechanistic interpretability for State-Space Models: SAEs, feature visualization, and a Hub registry for Mamba/Mamba-2.
+Project-URL: Homepage, https://github.com/hinanohart/recurrentlens
+Project-URL: Issues, https://github.com/hinanohart/recurrentlens/issues
+Author: hinanohart
+License-Expression: Apache-2.0
+License-File: LICENSE
+Keywords: interpretability,mamba,mechanistic-interpretability,sparse-autoencoder,ssm,state-space-model
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.11
+Requires-Dist: datasets>=2.20
+Requires-Dist: einops>=0.7
+Requires-Dist: huggingface-hub>=0.26
+Requires-Dist: jinja2>=3.1
+Requires-Dist: numpy>=1.26
+Requires-Dist: pydantic>=2.7
+Requires-Dist: pyyaml>=6.0
+Requires-Dist: safetensors>=0.4
+Requires-Dist: torch>=2.4
+Requires-Dist: tqdm>=4.66
+Requires-Dist: transformers>=4.45
+Requires-Dist: zarr<3.0,>=2.18
+Provides-Extra: dev
+Requires-Dist: ipykernel>=6.29; extra == 'dev'
+Requires-Dist: pyright>=1.1; extra == 'dev'
+Requires-Dist: pytest-cov>=5.0; extra == 'dev'
+Requires-Dist: pytest>=8.0; extra == 'dev'
+Requires-Dist: ruff>=0.6; extra == 'dev'
+Provides-Extra: mamba
+Requires-Dist: mamba-ssm>=2.2.2; extra == 'mamba'
+Description-Content-Type: text/markdown
+
+# recurrentlens
+
+**Mechanistic interpretability for State-Space Models.** Sparse autoencoders, feature visualization, and a Hub registry for Mamba / Mamba-2.
+
+[![CI](https://github.com/hinanohart/recurrentlens/actions/workflows/ci.yml/badge.svg)](https://github.com/hinanohart/recurrentlens/actions/workflows/ci.yml)
+[![License: Apache-2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
+[![Python 3.11+](https://img.shields.io/badge/python-3.11+-blue.svg)](https://www.python.org/downloads/)
+[![Status: Alpha](https://img.shields.io/badge/status-alpha-orange.svg)](#status)
+
+**Colab notebooks:** [Quickstart](https://colab.research.google.com/github/hinanohart/recurrentlens/blob/main/notebooks/01_quickstart.ipynb) · [Explore an SAE](https://colab.research.google.com/github/hinanohart/recurrentlens/blob/main/notebooks/02_explore_pretrained.ipynb) · [Train your own](https://colab.research.google.com/github/hinanohart/recurrentlens/blob/main/notebooks/03_train_mamba130m_sae.ipynb)
+
+---
+
+## Status
+
+**Alpha.** The public API is intentionally small (see [v0.1.0 API](#v010-api)) and may evolve through v0.1.x. Pretrained SAE artifacts are **not** yet on the Hub — they ship in v0.1.1 via a community Colab T4 run. See [CHANGELOG.md](./CHANGELOG.md) for the v0.1.0.post1 correctness hotfix (JumpReLU STE, HTML escape, CE metric cleanup).
+
+## ⚠️ v0.1.0.post1 scope disclosure (read first)
+
+`recurrentlens` ships a **framework + smoke-tested API + 3 Colab notebooks**. **Pretrained SAE artifacts for Mamba-130M/1.3B will land in v0.1.1**, trained via the included Colab notebook (T4 GPU is sufficient). The maintainer's build environment is CPU-only; the artifact training is deferred rather than fabricated.
+
+If you have a GPU and want to seed the Hub registry early, run `notebooks/03_train_mamba130m_sae.ipynb` and `recurrentlens.hub.push_sae(...)` — pull requests welcome.
+
+## Why recurrentlens
+
+SAELens is the canonical SAE training library and supports a wide range of architectures. As of 2026-05, the Mamba path has [open issues since 2024-10](https://github.com/decoderesearch/SAELens/issues/311), and there is room for an SSM-first design that treats recurrent state and selective scan as first-class concerns rather than retrofitting them onto a transformer-shaped abstraction. `recurrentlens` is that design.
+
+## Install
+
+```bash
+uv pip install recurrentlens                # core (CPU smoke + scaffold)
+uv pip install "recurrentlens[mamba]"        # adds mamba-ssm CUDA kernels — Linux + CUDA only
+uv pip install "recurrentlens[dev]"          # development
+```
+
+> The `[mamba]` extra pulls `mamba-ssm`, which requires CUDA and a Linux toolchain. CPU users and macOS users should install just `recurrentlens` (the core works via HF Transformers' Mamba implementation).
+
+## Quick look
+
+```python
+import recurrentlens as rl
+
+model = rl.load_model("state-spaces/mamba-130m-hf")
+sae = rl.train_sae(
+    model,
+    hook_site="out_proj_out",   # residual-stream analog (default)
+    layer=6,
+    variant="topk",
+    d_sae=16384,
+    n_tokens=200_000,            # CPU smoke; use 5M-100M+ on a T4 GPU
+)
+
+with rl.steer(model, sae, feature_id=42, vector_scale=2.0):
+    out = model.generate("The capital of France is", max_new_tokens=20)
+
+# Capture activations + contexts first; see notebooks/02 for a runnable example.
+# rl.viz.feature_explorer(sae, feature_id=42, activations=acts, contexts=ctxs).save("feature42.html")
+```
+
+> **`n_tokens` note.** 5M tokens on CPU takes hours; 200k completes in a couple of minutes for a smoke check. For research-grade SAEs the design target is 100M+ tokens, which is what the v0.1.1 pretrained artifacts will be trained on via Colab T4. See `notebooks/03_train_mamba130m_sae.ipynb`.
+
+See `notebooks/` for end-to-end examples.
+
+## v0.1.0 API
+
+| Function | Status |
+|---|---|
+| `load_model(model_id, device, dtype)` | scaffold (Phase 2) |
+| `train_sae(model, hook_site, layer, variant, ...)` | scaffold (Phases 4–5) |
+| `ablate(model, sae, feature_id, strength)` | scaffold (Phase 7) |
+| `steer(model, sae, feature_id, vector_scale)` | scaffold (Phase 7) |
+| `viz.feature_explorer(sae, feature_id, ...)` | scaffold (Phase 8) |
+| `hub.load_sae(repo_id)` / `hub.push_sae(sae, repo_id)` | scaffold (Phase 8) |
+| `bench.evaluate(sae, metrics=[...])` | scaffold (Phase 9) |
+
+`extract_circuit` and `diff.compare` are scheduled for **v0.2**.
+
+## Supported models (v0.1.0)
+
+- `state-spaces/mamba-130m-hf` (smoke + Colab training)
+- `state-spaces/mamba-130m` (raw)
+- `state-spaces/mamba2-1.3b` (Colab training only — too large for CI smoke)
+
+Falcon-Mamba-7B, RWKV-7, and Jamba are **v0.2+** targets.
+
+## Honest assessment
+
+- **Novelty**: As of 2026-05, OSS coverage of SSM-specific SAE training is thin; SAELens has Mamba code paths but [Issue #311](https://github.com/decoderesearch/SAELens/issues/311) has been open since 2024-10. `recurrentlens` treats the SSM design (recurrent state, selective scan, `out_proj` residual) as first-class.
+- **Practicality**: `load_model` + `train_sae` + `viz.feature_explorer` + `hub.{load,push}_sae` gives users a runnable scaffold today; with v0.1.1 pretrained artifacts the path becomes a 5-minute end-to-end demo.
+- **Scope**: v0.1.0.post1 has **no** pretrained SAE artifacts. Real recon-MSE / L0 / CE-recovery numbers will be published with v0.1.1 SAEs (trained via Colab T4). The shipped `train_sae` default is `n_tokens=200_000` for CPU smoke; the research-grade target is 100M+ tokens, which v0.1.1 will use.
+- **Open question**: do Mamba `out_proj_out` activations actually yield monosemantic SAE features? An empirical question the v0.1.1 release will start answering.
+
+## Pretrained SAEs
+
+| repo_id | model | layer | hook | variant | d_sae | status |
+|---|---|---|---|---|---|---|
+| `hinanohart/recurrentlens-mamba130m-L2-sae` | mamba-130m | 2 | out_proj_out | topk | 16384 | **planned v0.1.1** |
+| `hinanohart/recurrentlens-mamba130m-L6-sae` | mamba-130m | 6 | out_proj_out | topk | 16384 | **planned v0.1.1** |
+| `hinanohart/recurrentlens-mamba130m-L10-sae` | mamba-130m | 10 | out_proj_out | topk | 16384 | **planned v0.1.1** |
+
+Want to seed this table? Run notebook 03 on Colab T4 and open a PR. See [CONTRIBUTING.md](CONTRIBUTING.md).
+
+## Citation
+
+See `CITATION.cff`.
+
+## License
+
+Apache-2.0.