ablate-llm 0.2.0__tar.gz → 0.3.0__tar.gz

ablate_llm-0.3.0/LICENSE

@@ -0,0 +1,32 @@
+Proprietary License — All Rights Reserved
+
+Copyright (c) 2026 Sonu Kumar / AI Anytime
+
+This repository is publicly available for viewing and evaluation
+purposes only.
+
+No license is granted to use, reproduce, modify, distribute,
+sublicense, sell, commercially exploit, or incorporate this software,
+its documentation, or any substantial portion thereof into another
+product, service, dataset, or machine learning training corpus without
+prior written permission from the copyright holder.
+
+Viewing the source code on GitHub, and forking it within GitHub as
+permitted by the GitHub Terms of Service, does not grant any further
+rights.
+
+Third-party components included in this repository (for example,
+vendored libraries) remain subject to their own licenses.
+
+This license applies to all versions of this repository published on
+or after the date of this change. Versions previously published under
+a different license remain available only under the terms of that
+earlier license.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED. IN NO EVENT SHALL THE COPYRIGHT HOLDER BE LIABLE
+FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY ARISING FROM THE SOFTWARE.
+
+For commercial licensing inquiries:
+aianytime07@gmail.com
+sonu@aianytime.net

{ablate_llm-0.2.0 → ablate_llm-0.3.0}/PKG-INFO

@@ -1,9 +1,9 @@
 Metadata-Version: 2.4
 Name: ablate-llm
-Version: 0.2.0
+Version: 0.3.0
 Summary: Directional ablation (abliteration) toolkit for open-source language models: extract, ablate, and optimize the removal of refusal / behavioral directions in the residual stream.
 Author: AI Anytime
-License: MIT
+License: Proprietary
 Project-URL: Homepage, https://github.com/AIAnytime/ablate
 Project-URL: Repository, https://github.com/AIAnytime/ablate
 Project-URL: Issues, https://github.com/AIAnytime/ablate/issues
@@ -11,7 +11,7 @@ Project-URL: Example model, https://huggingface.co/ai-anytime/qwen-1.5b-ablitera
 Keywords: interpretability,activation-engineering,abliteration,refusal-direction,mechanistic-interpretability,safety
 Classifier: Development Status :: 4 - Beta
 Classifier: Intended Audience :: Science/Research
-Classifier: License :: OSI Approved :: MIT License
+Classifier: License :: Other/Proprietary License
 Classifier: Programming Language :: Python :: 3
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.9
@@ -55,18 +55,25 @@ Dynamic: license-file
 <p align="center">
   <a href="https://pypi.org/project/ablate-llm/"><img alt="PyPI" src="https://img.shields.io/pypi/v/ablate-llm.svg?color=7c5cff"></a>
   <a href="https://pypi.org/project/ablate-llm/"><img alt="Python" src="https://img.shields.io/pypi/pyversions/ablate-llm.svg"></a>
-  <a href="https://github.com/AIAnytime/ablate/blob/main/LICENSE"><img alt="License: MIT" src="https://img.shields.io/badge/license-MIT-green.svg"></a>
+  <a href="https://github.com/AIAnytime/ablate/blob/main/LICENSE"><img alt="License: Proprietary" src="https://img.shields.io/badge/license-Proprietary-red.svg"></a>
 </p>
 
 <p align="center">
   <a href="https://pypi.org/project/ablate-llm/">📦 PyPI</a> &nbsp;•&nbsp;
   <a href="https://huggingface.co/ai-anytime/qwen-1.5b-abliterated">🤗 Model</a> &nbsp;•&nbsp;
-  <a href="https://www.youtube.com/@AIAnytime">▶️ Video walkthrough (coming soon)</a> &nbsp;•&nbsp;
+  <a href="https://youtu.be/1-1bM3Ynql4">▶️ Video walkthrough</a> &nbsp;•&nbsp;
   <a href="#citation">📚 Cite</a>
 </p>
 
 <p align="center"><i>Created with 💜 by <b>AI Anytime</b></i></p>
 
+<p align="center">
+  <a href="https://youtu.be/1-1bM3Ynql4">
+    <img alt="Watch the video walkthrough" src="https://img.youtube.com/vi/1-1bM3Ynql4/maxresdefault.jpg" width="560">
+  </a>
+  <br><i>▶️ Watch the full walkthrough on YouTube</i>
+</p>
+
 ---
 
 `ablate` finds the linear direction in a transformer's residual stream that
@@ -270,6 +277,35 @@ ablate extract --model gpt2 --method diff_of_means --output directions.pt
 
 ---
 
+## Browser UI
+
+```bash
+ablate ui                                   # opens http://localhost:7860
+ablate ui --model Qwen/Qwen2.5-0.5B-Instruct  # preload a model
+ablate ui --share                           # let others on the same Wi-Fi explore
+```
+
+A local, dependency-free web app that walks through the whole pipeline:
+
+1. **Model** — search the HuggingFace Hub and load any chat model.
+2. **Find direction** — extract the refusal direction and see, per layer, how
+   well it separates the two prompt piles and how much refusal is left when it is
+   removed. Every prompt is plotted along the direction.
+3. **Playground** — the original and the edited model answer the same prompt
+   side by side, streamed from one forward pass. Includes the first-token
+   probabilities, a per-token/per-layer "X-ray" of the refusal signal, and a
+   *Force refusal* mode that adds the direction instead of removing it.
+4. **Measure** — refusal rate before/after on held-out prompts plus KL drift.
+5. **Auto-tune** — the Optuna search, plotted live.
+6. **Publish** — bake and push to the Hub (private by default) with the
+   generated model card, or save locally. Your HF token is picked up from
+   `HF_TOKEN`, a local `.env`, or `huggingface-cli login`.
+7. **Architecture** and **How it works** — the system diagram and the method in
+   four pictures, for explaining it to someone else.
+
+With `--share`, other machines can use the playground and charts; loading
+models, saving and publishing stay restricted to the machine running the server.
+
 ## Colab notebooks
 
 Both install `ablate` from an uploaded `ablate-tool.zip` (no PyPI needed) and run
@@ -380,10 +416,11 @@ Is Mediated by a Single Direction*.
 
 ## Acknowledgements
 
-Created with 💜 by **[AI Anytime](https://www.youtube.com/@AIAnytime)**. A full video
-walkthrough is coming soon — subscribe on YouTube. Built on the mechanistic-interpretability
-work of Arditi et al. and the open-source `transformers` / `optuna` ecosystems.
+Created with 💜 by **[AI Anytime](https://www.youtube.com/@AIAnytime)**. Watch the full
+**[video walkthrough on YouTube](https://youtu.be/1-1bM3Ynql4)**. Built on the
+mechanistic-interpretability work of Arditi et al. and the open-source
+`transformers` / `optuna` ecosystems.
 
 ## License
 
-MIT © AI Anytime.
+Proprietary — all rights reserved. This code is published for viewing and evaluation only; no use, copying, modification, redistribution, commercial use, or use as AI/ML training data without written permission. See [LICENSE](LICENSE). Commercial licensing: aianytime07@gmail.com · sonu@aianytime.net.

{ablate_llm-0.2.0 → ablate_llm-0.3.0}/README.md

@@ -13,18 +13,25 @@
 <p align="center">
   <a href="https://pypi.org/project/ablate-llm/"><img alt="PyPI" src="https://img.shields.io/pypi/v/ablate-llm.svg?color=7c5cff"></a>
   <a href="https://pypi.org/project/ablate-llm/"><img alt="Python" src="https://img.shields.io/pypi/pyversions/ablate-llm.svg"></a>
-  <a href="https://github.com/AIAnytime/ablate/blob/main/LICENSE"><img alt="License: MIT" src="https://img.shields.io/badge/license-MIT-green.svg"></a>
+  <a href="https://github.com/AIAnytime/ablate/blob/main/LICENSE"><img alt="License: Proprietary" src="https://img.shields.io/badge/license-Proprietary-red.svg"></a>
 </p>
 
 <p align="center">
   <a href="https://pypi.org/project/ablate-llm/">📦 PyPI</a> &nbsp;•&nbsp;
   <a href="https://huggingface.co/ai-anytime/qwen-1.5b-abliterated">🤗 Model</a> &nbsp;•&nbsp;
-  <a href="https://www.youtube.com/@AIAnytime">▶️ Video walkthrough (coming soon)</a> &nbsp;•&nbsp;
+  <a href="https://youtu.be/1-1bM3Ynql4">▶️ Video walkthrough</a> &nbsp;•&nbsp;
   <a href="#citation">📚 Cite</a>
 </p>
 
 <p align="center"><i>Created with 💜 by <b>AI Anytime</b></i></p>
 
+<p align="center">
+  <a href="https://youtu.be/1-1bM3Ynql4">
+    <img alt="Watch the video walkthrough" src="https://img.youtube.com/vi/1-1bM3Ynql4/maxresdefault.jpg" width="560">
+  </a>
+  <br><i>▶️ Watch the full walkthrough on YouTube</i>
+</p>
+
 ---
 
 `ablate` finds the linear direction in a transformer's residual stream that
@@ -228,6 +235,35 @@ ablate extract --model gpt2 --method diff_of_means --output directions.pt
 
 ---
 
+## Browser UI
+
+```bash
+ablate ui                                   # opens http://localhost:7860
+ablate ui --model Qwen/Qwen2.5-0.5B-Instruct  # preload a model
+ablate ui --share                           # let others on the same Wi-Fi explore
+```
+
+A local, dependency-free web app that walks through the whole pipeline:
+
+1. **Model** — search the HuggingFace Hub and load any chat model.
+2. **Find direction** — extract the refusal direction and see, per layer, how
+   well it separates the two prompt piles and how much refusal is left when it is
+   removed. Every prompt is plotted along the direction.
+3. **Playground** — the original and the edited model answer the same prompt
+   side by side, streamed from one forward pass. Includes the first-token
+   probabilities, a per-token/per-layer "X-ray" of the refusal signal, and a
+   *Force refusal* mode that adds the direction instead of removing it.
+4. **Measure** — refusal rate before/after on held-out prompts plus KL drift.
+5. **Auto-tune** — the Optuna search, plotted live.
+6. **Publish** — bake and push to the Hub (private by default) with the
+   generated model card, or save locally. Your HF token is picked up from
+   `HF_TOKEN`, a local `.env`, or `huggingface-cli login`.
+7. **Architecture** and **How it works** — the system diagram and the method in
+   four pictures, for explaining it to someone else.
+
+With `--share`, other machines can use the playground and charts; loading
+models, saving and publishing stay restricted to the machine running the server.
+
 ## Colab notebooks
 
 Both install `ablate` from an uploaded `ablate-tool.zip` (no PyPI needed) and run
@@ -338,10 +374,11 @@ Is Mediated by a Single Direction*.
 
 ## Acknowledgements
 
-Created with 💜 by **[AI Anytime](https://www.youtube.com/@AIAnytime)**. A full video
-walkthrough is coming soon — subscribe on YouTube. Built on the mechanistic-interpretability
-work of Arditi et al. and the open-source `transformers` / `optuna` ecosystems.
+Created with 💜 by **[AI Anytime](https://www.youtube.com/@AIAnytime)**. Watch the full
+**[video walkthrough on YouTube](https://youtu.be/1-1bM3Ynql4)**. Built on the
+mechanistic-interpretability work of Arditi et al. and the open-source
+`transformers` / `optuna` ecosystems.
 
 ## License
 
-MIT © AI Anytime.
+Proprietary — all rights reserved. This code is published for viewing and evaluation only; no use, copying, modification, redistribution, commercial use, or use as AI/ML training data without written permission. See [LICENSE](LICENSE). Commercial licensing: aianytime07@gmail.com · sonu@aianytime.net.

{ablate_llm-0.2.0 → ablate_llm-0.3.0}/pyproject.toml

@@ -4,11 +4,11 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "ablate-llm"
-version = "0.2.0"
+version = "0.3.0"
 description = "Directional ablation (abliteration) toolkit for open-source language models: extract, ablate, and optimize the removal of refusal / behavioral directions in the residual stream."
 readme = "README.md"
 requires-python = ">=3.9"
-license = { text = "MIT" }
+license = { text = "Proprietary" }
 authors = [{ name = "AI Anytime" }]
 keywords = [
     "interpretability",
@@ -21,7 +21,7 @@ keywords = [
 classifiers = [
     "Development Status :: 4 - Beta",
     "Intended Audience :: Science/Research",
-    "License :: OSI Approved :: MIT License",
+    "License :: Other/Proprietary License",
     "Programming Language :: Python :: 3",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
@@ -52,4 +52,4 @@ ablate = "ablate.cli:main"
 where = ["src"]
 
 [tool.setuptools.package-data]
-ablate = ["_data/*.txt"]
+ablate = ["_data/*.txt", "ui/static/*"]

{ablate_llm-0.2.0 → ablate_llm-0.3.0}/src/ablate/__init__.py

@@ -23,7 +23,7 @@ from .pipeline import Ablator, run
 from .weights import bake_direction, bake_subspace
 from .hub import build_model_card, push_to_hub
 
-__version__ = "0.2.0"
+__version__ = "0.3.0"
 
 __all__ = [
     "Ablator",

{ablate_llm-0.2.0 → ablate_llm-0.3.0}/src/ablate/cli.py

@@ -1,4 +1,4 @@
-"""Command-line interface: ``ablate run|extract|generate|eval``."""
+"""Command-line interface: ``ablate run|extract|generate|eval|ui``."""
 from __future__ import annotations
 
 import argparse
@@ -104,6 +104,14 @@ def cmd_eval(args: argparse.Namespace) -> int:
     return 0
 
 
+def cmd_ui(args: argparse.Namespace) -> int:
+    from .ui import serve
+
+    serve(host="0.0.0.0" if args.share else args.host, port=args.port, model=args.model,
+          open_browser=not args.no_browser)
+    return 0
+
+
 def build_parser() -> argparse.ArgumentParser:
     p = argparse.ArgumentParser(prog="ablate", description="Directional ablation toolkit for language models.")
     sub = p.add_subparsers(dest="command", required=True)
@@ -158,6 +166,15 @@ def build_parser() -> argparse.ArgumentParser:
     ev.add_argument("--max-new-tokens", type=int, default=128)
     ev.set_defaults(func=cmd_eval)
 
+    u = sub.add_parser("ui", help="open the browser UI (explore, ablate, measure, publish)")
+    u.add_argument("--model", default=None, help="model to preload (otherwise pick one in the UI)")
+    u.add_argument("--host", default="127.0.0.1")
+    u.add_argument("--port", type=int, default=7860)
+    u.add_argument("--share", action="store_true",
+                   help="listen on the local network so others on the same Wi-Fi can explore")
+    u.add_argument("--no-browser", action="store_true", help="do not open a browser tab")
+    u.set_defaults(func=cmd_ui)
+
     return p
 
 

{ablate_llm-0.2.0 → ablate_llm-0.3.0}/src/ablate/hub.py

@@ -59,7 +59,7 @@ def build_model_card(
 
 This is a **directionally-ablated ("abliterated")** version of
 [`{base_model}`](https://huggingface.co/{base_model}), produced with
-[**ablate**](https://github.com/) — an activation-engineering toolkit for
+[**ablate**](https://github.com/AIAnytime/ablate) — an activation-engineering toolkit for
 mechanistic-interpretability and safety research.
 
 The model's **refusal direction(s)** were identified in the residual stream and

{ablate_llm-0.2.0 → ablate_llm-0.3.0}/src/ablate/optimize.py

@@ -11,7 +11,7 @@ guards against degenerate solutions that "win" by breaking the model.
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import List, Optional
+from typing import Callable, List, Optional
 
 import torch
 
@@ -47,6 +47,7 @@ def optimize(
     layer_lo: Optional[int] = None,
     layer_hi: Optional[int] = None,
     show_progress: bool = True,
+    callback: Optional[Callable[[dict], None]] = None,
 ) -> SearchResult:
     try:
         import optuna
@@ -92,6 +93,8 @@ def optimize(
         penalty = 0.0 if res.coherence >= coherence_floor else (coherence_floor - res.coherence) * 5.0
         score = res.objective + penalty
         history.append({"trial": trial.number, **cfg.to_dict(), **res.__dict__, "score": score})
+        if callback is not None:
+            callback(history[-1])
         return score
 
     sampler = optuna.samplers.TPESampler(seed=seed)
@@ -129,6 +132,7 @@ def optimize_subspace(
     coherence_floor: float = 0.3,
     seed: int = 0,
     show_progress: bool = True,
+    callback: Optional[Callable[[dict], None]] = None,
 ) -> SubspaceSearchResult:
     """Search over ``(n_directions, alpha, layer_band)`` for subspace ablation.
 
@@ -161,6 +165,8 @@ def optimize_subspace(
         penalty = 0.0 if res.coherence >= coherence_floor else (coherence_floor - res.coherence) * 5.0
         score = res.objective + penalty
         history.append({"trial": trial.number, **cfg.to_dict(), **res.__dict__, "score": score})
+        if callback is not None:
+            callback(history[-1])
         return score
 
     study = optuna.create_study(direction="minimize", sampler=optuna.samplers.TPESampler(seed=seed))

ablate_llm-0.3.0/src/ablate/ui/__init__.py

@@ -0,0 +1,4 @@
+"""Browser UI for ablate — launch with ``ablate ui``."""
+from .server import serve
+
+__all__ = ["serve"]