mi-crow 0.1.1.post12__tar.gz → 0.1.1.post14__tar.gz

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/.cursor/plans/dokumentacja-implementacji-modu-w-i-automatyzacji-83a10087.plan.md

@@ -69,14 +69,14 @@ Jeden plik Markdown (np. `raport.md` lub `raport_implementacji.md`) w katalogu g
 
 ## 2. Implementacja modułów
 
-### 2.1 Moduł `amber.datasets`
+### 2.1 Moduł `mi_crow.datasets`
 
 - **BaseDataset**: Klasa abstrakcyjna z LoadingStrategy (MEMORY, DYNAMIC_LOAD, ITERABLE_ONLY)
 - **TextDataset**: Dataset tekstowy z parametrem `text_field`
 - **ClassificationDataset**: Dataset z kategoriami, parametry `text_field`, `category_field` (single/multiple)
 - **LoadingStrategy**: Enum z trzema strategiami ładowania
 
-### 2.2 Moduł `amber.hooks`
+### 2.2 Moduł `mi_crow.hooks`
 
 - **Hook**: Klasa abstrakcyjna, parametry: `layer_signature`, `hook_type` (FORWARD/PRE_FORWARD), `hook_id`
 - **Detector**: Hook do wykrywania/zapisywania aktywacji
@@ -84,7 +84,7 @@ Jeden plik Markdown (np. `raport.md` lub `raport_implementacji.md`) w katalogu g
 - **LayerActivationDetector**: Implementacja zapisywania aktywacji warstw
 - **FunctionController**: Implementacja kontrolera funkcji
 
-### 2.3 Moduł `amber.language_model`
+### 2.3 Moduł `mi_crow.language_model`
 
 - **LanguageModel**: Główna klasa wrappera modelu
 - Parametry inicjalizacji: `model`, `tokenizer`, `store`, `model_id`
@@ -95,7 +95,7 @@ Jeden plik Markdown (np. `raport.md` lub `raport_implementacji.md`) w katalogu g
 - **LanguageModelTokenizer**: Wrapper tokenizera
 - **InferenceEngine**: Silnik inferencji z parametrami `autocast`, `autocast_dtype`, `with_controllers`
 
-### 2.4 Moduł `amber.mechanistic.sae` (moduły uczenia maszynowego)
+### 2.4 Moduł `mi_crow.mechanistic.sae` (moduły uczenia maszynowego)
 
 - **Sae**: Klasa abstrakcyjna SAE
 - Parametry: `n_latents`, `n_inputs`, `hook_id`, `device`, `store`
@@ -111,7 +111,7 @@ Jeden plik Markdown (np. `raport.md` lub `raport_implementacji.md`) w katalogu g
 - **AutoencoderContext**: Kontekst SAE z parametrami: `n_latents`, `n_inputs`, `device`, `text_tracking_enabled`, `text_tracking_k`
 - **AutoencoderConcepts**: Zarządzanie konceptami (multiplication, bias)
 
-### 2.5 Moduł `amber.store`
+### 2.5 Moduł `mi_crow.store`
 
 - **Store**: Klasa abstrakcyjna do przechowywania tensorów
 - Parametry: `base_path`, `runs_prefix`, `dataset_prefix`, `model_prefix`

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/.cursor/plans/server-sae-full-metadata_b869cacf.plan.md

@@ -1,6 +1,6 @@
 ---
 name: server-sae-full-metadata
-overview: "Wire server to amber SAE without touching core library: LocalStore-based activations, real SAE training/inference, concept application, and enriched responses."
+overview: "Wire server to mi_crow SAE without touching core library: LocalStore-based activations, real SAE training/inference, concept application, and enriched responses."
 todos: []
 ---
 
@@ -10,21 +10,21 @@ todos: []
 
 b
 
-- Only modify server code; reuse amber APIs as-is.
+- Only modify server code; reuse mi_crow APIs as-is.
 - Use LocalStore for activations, real SaeTrainer for training, and Sae hook for inference/top-texts.
-- Concept handling via existing amber concept classes; no edits to amber package.
+- Concept handling via existing mi_crow concept classes; no edits to mi_crow package.
 
 ## Plan
 
 1) **Activation Pipeline → LocalStore**
 
-- Update `/sae/activations/save` to write activations into `amber.store.LocalStore` under `artifact_base_path` using `run_id`/batch/layer keys.
+- Update `/sae/activations/save` to write activations into `mi_crow.store.LocalStore` under `artifact_base_path` using `run_id`/batch/layer keys.
 - Keep manifest with run_id, layers, dataset info, token counts, batch indices.
 - Add layer-existence check (already present) and bounds on batch/sample size.
 
 2) **SAE Training (real)**
 
-- In `/sae/train`, load LocalStore with run_id/layer from manifest; build `SaeTrainingConfig` from payload; call `SaeTrainer.train` on the selected SAE class (chosen by payload or default) without modifying amber code.
+- In `/sae/train`, load LocalStore with run_id/layer from manifest; build `SaeTrainingConfig` from payload; call `SaeTrainer.train` on the selected SAE class (chosen by payload or default) without modifying mi_crow code.
 - Save SAE checkpoint via its `save` method alongside training metadata (history, config, metrics) in `artifact_base_path/sae/<model>/<run>/`.
 - Return sae_id/sae_path/metadata_path from the job result.
 
@@ -36,9 +36,9 @@ b
 
 4) **Concepts Integration**
 
-- `/sae/concepts/load`: parse user file into amber concepts/dictionary; validate against SAE dims; persist normalized concept file under concepts dir.
+- `/sae/concepts/load`: parse user file into mi_crow concepts/dictionary; validate against SAE dims; persist normalized concept file under concepts dir.
 - `/sae/concepts/manipulate`: accept weights/edits; build a concept-config mapping to neuron weights; persist config.
-- `/sae/infer` accepts concept_config to adjust activations/decoder via SAE hook before generation (no amber code changes—use available hook points).
+- `/sae/infer` accepts concept_config to adjust activations/decoder via SAE hook before generation (no mi_crow code changes—use available hook points).
 
 5) **Jobs & Observability**
 
@@ -46,10 +46,10 @@ b
 
 6) **Testing & Docs**
 
-- Extend server tests with small dummy SAE class implementing `save/load/encode/decode` to cover training/inference flows without touching amber; mock LocalStore I/O.
+- Extend server tests with small dummy SAE class implementing `save/load/encode/decode` to cover training/inference flows without touching mi_crow; mock LocalStore I/O.
 - Update README with new expectations: activations in LocalStore, real training path, concept flow, top-text outputs, and job responses.
 
 ## Notes
 
-- No changes to amber core; all integrations via its public classes (LocalStore, SaeTrainer, Sae, concepts).
+- No changes to mi_crow core; all integrations via its public classes (LocalStore, SaeTrainer, Sae, concepts).
 - Keep payload schemas mostly stable; only enrich responses with sae/metadata paths and top-text summaries.

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/.github/workflows/tests.yml

@@ -34,9 +34,10 @@ jobs:
 
       - name: Verify package installation
         run: |
-          python -c "import amber; print('✓ amber package found')"
-          python -c "import amber.store; print('✓ amber.store package found')"
-          python -c "from amber.store.store import Store; print('✓ Store import successful')"
+          python -c "import mi_crow; print('✓ mi_crow package found')"
+          python -c "import sys; import os; import mi_crow; print('mi_crow location:', mi_crow.__file__); print('mi_crow.__path__:', getattr(mi_crow, '__path__', 'NOT SET')); store_dir = os.path.join(os.path.dirname(mi_crow.__file__), 'store'); print('store dir exists:', os.path.exists(store_dir)); print('store/__init__.py exists:', os.path.exists(os.path.join(store_dir, '__init__.py')))"
+          python -c "import mi_crow.store; print('✓ mi_crow.store package found')"
+          python -c "from mi_crow.store.store import Store; print('✓ Store import successful')"
 
       - name: Run tests with coverage threshold
         run: |

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/.gitignore

@@ -92,7 +92,8 @@ outputs
 # experiments
 .llm_context/
 results/
-store/
+# Exclude data store directories but NOT src/mi_crow/store/ (package code)
+/experiments/*/store/
 
 # frontend
 src/frontend/node_modules/
@@ -100,4 +101,4 @@ src/frontend/dist/
 src/frontend/build/
 src/frontend/.next/
 
-amber_server
+mi_crow_server

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mi-crow
-Version: 0.1.1.post12
+Version: 0.1.1.post14
 Summary: Engineer Thesis: Explaining and modifying LLM responses using SAE and concepts.
 Author-email: Hubert Kowalski <your.email@example.com>, Adam Kaniasty <adam.kaniasty@gmail.com>
 Requires-Python: >=3.10
@@ -96,7 +96,7 @@ uv run --group server pytest tests/server/test_api.py --cov=server --cov-fail-un
 
 ### SAE API usage
 
-- Configure artifact location (optional): `export SERVER_ARTIFACT_BASE_PATH=/path/to/amber_artifacts` (defaults to `~/.cache/amber_server`)
+- Configure artifact location (optional): `export SERVER_ARTIFACT_BASE_PATH=/path/to/mi_crow_artifacts` (defaults to `~/.cache/mi_crow_server`)
 - Load a model: `curl -X POST http://localhost:8000/models/load -H "Content-Type: application/json" -d '{"model_id":"bielik"}'`
 - Save activations from dataset (stored in `LocalStore` under `activations/<model>/<run_id>`):
   - HF dataset: `{"dataset":{"type":"hf","name":"ag_news","split":"train","text_field":"text"}}`

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/README.md

@@ -58,7 +58,7 @@ uv run --group server pytest tests/server/test_api.py --cov=server --cov-fail-un
 
 ### SAE API usage
 
-- Configure artifact location (optional): `export SERVER_ARTIFACT_BASE_PATH=/path/to/amber_artifacts` (defaults to `~/.cache/amber_server`)
+- Configure artifact location (optional): `export SERVER_ARTIFACT_BASE_PATH=/path/to/mi_crow_artifacts` (defaults to `~/.cache/mi_crow_server`)
 - Load a model: `curl -X POST http://localhost:8000/models/load -H "Content-Type: application/json" -d '{"model_id":"bielik"}'`
 - Save activations from dataset (stored in `LocalStore` under `activations/<model>/<run_id>`):
   - HF dataset: `{"dataset":{"type":"hf","name":"ag_news","split":"train","text_field":"text"}}`

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/docs/api/datasets.md

@@ -2,5 +2,5 @@
 
 Dataset loading and management utilities for text and classification datasets.
 
-::: amber.datasets
+::: mi_crow.datasets
 

mi_crow-0.1.1.post14/docs/api/hooks.md

@@ -0,0 +1,26 @@
+# Hooks API
+
+Hook system for intercepting and managing model activations during inference.
+
+## Core Hook Classes
+
+::: mi_crow.hooks.hook.Hook
+
+::: mi_crow.hooks.detector.Detector
+
+::: mi_crow.hooks.controller.Controller
+
+## Implementations
+
+::: mi_crow.hooks.implementations.layer_activation_detector.LayerActivationDetector
+
+::: mi_crow.hooks.implementations.model_input_detector.ModelInputDetector
+
+::: mi_crow.hooks.implementations.model_output_detector.ModelOutputDetector
+
+::: mi_crow.hooks.implementations.function_controller.FunctionController
+
+## Utilities
+
+::: mi_crow.hooks.utils
+

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/docs/api/index.md

@@ -1,8 +1,8 @@
 # API Reference
 
-Amber's public Python API is documented automatically from docstrings.
+mi_crow's public Python API is documented automatically from docstrings.
 
-The top-level `amber` package is intentionally minimal (it only exports things
+The top-level `mi_crow` package is intentionally minimal (it only exports things
 like `ping`). The real functionality lives in subpackages, which are documented
 in the sections below.
 
@@ -16,5 +16,5 @@ in the sections below.
 
 ## Top-level Package
 
-::: amber
+::: mi_crow
 

mi_crow-0.1.1.post14/docs/api/language_model.md

@@ -0,0 +1,24 @@
+# Language Model API
+
+Core language model functionality for loading models, running inference, and managing activations.
+
+## Main Classes
+
+::: mi_crow.language_model.language_model.LanguageModel
+
+::: mi_crow.language_model.context.LanguageModelContext
+
+::: mi_crow.language_model.layers.LanguageModelLayers
+
+::: mi_crow.language_model.tokenizer.LanguageModelTokenizer
+
+::: mi_crow.language_model.activations.LanguageModelActivations
+
+::: mi_crow.language_model.inference.InferenceEngine
+
+## Utilities
+
+::: mi_crow.language_model.initialization
+
+::: mi_crow.language_model.persistence
+

mi_crow-0.1.1.post14/docs/api/sae.md

@@ -0,0 +1,34 @@
+# Sparse Autoencoder (SAE) API
+
+Sparse Autoencoders, training, concepts, and related modules for mechanistic interpretability.
+
+## Core SAE Classes
+
+::: mi_crow.mechanistic.sae.sae.Sae
+
+::: mi_crow.mechanistic.sae.modules.topk_sae.TopKSae
+
+::: mi_crow.mechanistic.sae.autoencoder_context.AutoencoderContext
+
+## Training
+
+::: mi_crow.mechanistic.sae.sae_trainer.SaeTrainer
+
+::: mi_crow.mechanistic.sae.sae_trainer.SaeTrainingConfig
+
+::: mi_crow.mechanistic.sae.modules.topk_sae.TopKSaeTrainingConfig
+
+## Concepts
+
+::: mi_crow.mechanistic.sae.concepts.autoencoder_concepts.AutoencoderConcepts
+
+::: mi_crow.mechanistic.sae.concepts.concept_dictionary.ConceptDictionary
+
+::: mi_crow.mechanistic.sae.concepts.concept_models
+
+::: mi_crow.mechanistic.sae.concepts.input_tracker.InputTracker
+
+## Training Utilities
+
+::: mi_crow.mechanistic.sae.training.wandb_logger.WandbLogger
+

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/docs/api/store.md

@@ -2,5 +2,5 @@
 
 Persistence layer for activations, models, and runs.
 
-::: amber.store
+::: mi_crow.store
 

mi_crow-0.1.1.post14/docs/api.md

@@ -0,0 +1,7 @@
+## API Reference
+
+mi_crow's public Python API is documented automatically from docstrings
+
+::: mi_crow
+
+

mi_crow-0.1.1.post14/docs/index.md

@@ -0,0 +1,5 @@
+# mi_crow
+
+mi_crow is a Python package for explaining and steering LLM behavior (SAE, concepts).
+
+- **Repo**: [GitHub](https://github.com/AdamKaniasty/Inzynierka)

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/mkdocs.yml

@@ -1,4 +1,4 @@
-site_name: Amber
+site_name: mi_crow
 site_description: Explaining and modifying LLM responses using SAE and concepts
 site_url: https://adamkaniasty.github.io/Inzynierka
 repo_url: https://github.com/AdamKaniasty/Inzynierka

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/pyproject.toml

@@ -50,13 +50,13 @@ docs = [
 ]
 
 [tool.pytest.ini_options]
-addopts = "--maxfail=1 -q --cov=amber --cov-report=term-missing --cov-report=xml --cov-fail-under=85"
+addopts = "--maxfail=1 -q --cov=mi_crow --cov-report=term-missing --cov-report=xml --cov-fail-under=85"
 testpaths = ["tests"]
 pythonpath = ["src"]
 
 [tool.coverage.run]
 branch = true
-source = ["amber"]
+source = ["mi_crow"]
 parallel = true
 
 [tool.coverage.report]
@@ -72,7 +72,7 @@ directory = "htmlcov"
 
 [tool.setuptools.packages.find]
 where = ["src"]
-include = ["amber*"]
+include = ["mi_crow*"]
 exclude = ["server*", "frontend*"]
 
 [tool.setuptools_scm]

{mi_crow-0.1.1.post12 → mi_crow-0.1.1.post14}/raport_implementacji.md

@@ -45,7 +45,7 @@
 ### 1.4 Testy i coverage (pytest)
 
 **Konfiguracja**: Sekcja `[tool.pytest.ini_options]` w `pyproject.toml`:
-- `addopts = "--maxfail=1 -q --cov=amber --cov-report=term-missing --cov-report=xml --cov-fail-under=90"`
+- `addopts = "--maxfail=1 -q --cov=mi_crow --cov-report=term-missing --cov-report=xml --cov-fail-under=90"`
 - `testpaths = ["tests"]`
 - `pythonpath = ["src"]`
 
@@ -96,9 +96,9 @@
 
 ## 2. Implementacja modułów
 
-### 2.1 Moduł amber.datasets
+### 2.1 Moduł mi_crow.datasets
 
-**BaseDataset** (`src/amber/datasets/base_dataset.py`):
+**BaseDataset** (`src/mi_crow/datasets/base_dataset.py`):
 - Klasa abstrakcyjna bazowa dla wszystkich datasetów
 - Parametry inicjalizacji:
   - `ds`: `Dataset | IterableDataset` - dataset HuggingFace
@@ -106,13 +106,13 @@
   - `loading_strategy`: `LoadingStrategy` - strategia ładowania (domyślnie `MEMORY`)
 - Metody abstrakcyjne: `__getitem__()`, `__len__()`, `__iter__()`
 
-**LoadingStrategy** (`src/amber/datasets/loading_strategy.py`):
+**LoadingStrategy** (`src/mi_crow/datasets/loading_strategy.py`):
 - Enum z trzema strategiami:
   - `MEMORY`: Ładowanie całego datasetu do pamięci (najszybszy dostęp, najwyższe użycie pamięci)
   - `DYNAMIC_LOAD`: Zapis na dysk, odczyt dynamiczny przez memory-mapped Arrow files (obsługuje len/getitem, niższe użycie pamięci)
   - `ITERABLE_ONLY`: Prawdziwy streaming używając IterableDataset (najniższe użycie pamięci, brak wsparcia dla len/getitem)
 
-**TextDataset** (`src/amber/datasets/text_dataset.py`):
+**TextDataset** (`src/mi_crow/datasets/text_dataset.py`):
 - Dataset tekstowy dziedziczący z `BaseDataset`
 - Parametry:
   - `ds`: `Dataset | IterableDataset`
@@ -120,7 +120,7 @@
   - `loading_strategy`: `LoadingStrategy` (domyślnie `MEMORY`)
   - `text_field`: `str` - nazwa kolumny z tekstem (domyślnie `"text"`)
 
-**ClassificationDataset** (`src/amber/datasets/classification_dataset.py`):
+**ClassificationDataset** (`src/mi_crow/datasets/classification_dataset.py`):
 - Dataset z kategoriami dziedziczący z `BaseDataset`
 - Parametry:
   - `ds`: `Dataset | IterableDataset`
@@ -129,9 +129,9 @@
   - `text_field`: `str` - nazwa kolumny z tekstem (domyślnie `"text"`)
   - `category_field`: `Union[str, List[str]]` - nazwa kolumny(ów) z kategorią/etykietą (domyślnie `"category"`), obsługuje pojedyncze lub wielokrotne etykiety
 
-### 2.2 Moduł amber.hooks
+### 2.2 Moduł mi_crow.hooks
 
-**Hook** (`src/amber/hooks/hook.py`):
+**Hook** (`src/mi_crow/hooks/hook.py`):
 - Klasa abstrakcyjna bazowa dla wszystkich hooków
 - Parametry inicjalizacji:
   - `layer_signature`: `str | int | None` - nazwa lub indeks warstwy do podpięcia hooka
@@ -139,7 +139,7 @@
   - `hook_id`: `str | None` - unikalny identyfikator (auto-generowany jeśli nie podany)
 - `HookType` Enum: `FORWARD`, `PRE_FORWARD`
 
-**Detector** (`src/amber/hooks/detector.py`):
+**Detector** (`src/mi_crow/hooks/detector.py`):
 - Klasa abstrakcyjna dziedzicząca z `Hook` do wykrywania/zapisywania aktywacji
 - Parametry:
   - `hook_type`: `HookType | str` (domyślnie `HookType.FORWARD`)
@@ -148,7 +148,7 @@
   - `layer_signature`: `str | int | None`
 - Atrybuty: `metadata: Dict[str, Any]`, `tensor_metadata: Dict[str, torch.Tensor]`
 
-**Controller** (`src/amber/hooks/controller.py`):
+**Controller** (`src/mi_crow/hooks/controller.py`):
 - Klasa abstrakcyjna dziedzicząca z `Hook` do modyfikacji aktywacji
 - Parametry:
   - `hook_type`: `HookType | str` (domyślnie `HookType.FORWARD`)
@@ -156,17 +156,17 @@
   - `layer_signature`: `str | int | None`
 - Może modyfikować wejścia (pre_forward) lub wyjścia (forward) warstw
 
-**LayerActivationDetector** (`src/amber/hooks/implementations/activation_saver.py`):
+**LayerActivationDetector** (`src/mi_crow/hooks/implementations/activation_saver.py`):
 - Implementacja `Detector` do zapisywania aktywacji warstw
 - Zapisuje aktywacje do Store podczas forward pass
 
-**FunctionController** (`src/amber/hooks/implementations/function_controller.py`):
+**FunctionController** (`src/mi_crow/hooks/implementations/function_controller.py`):
 - Implementacja `Controller` do kontroli funkcji
 - Pozwala na modyfikację aktywacji przez funkcję użytkownika
 
-### 2.3 Moduł amber.language_model
+### 2.3 Moduł mi_crow.language_model
 
-**LanguageModel** (`src/amber/language_model/language_model.py`):
+**LanguageModel** (`src/mi_crow/language_model/language_model.py`):
 - Główna klasa wrappera modelu językowego
 - Parametry inicjalizacji:
   - `model`: `nn.Module` - model PyTorch
@@ -183,19 +183,19 @@
   - `from_local_torch(model_path, tokenizer_path, store)` - ładowanie z lokalnych ścieżek HuggingFace
   - `from_local(saved_path, store, model_id)` - ładowanie z zapisanego pliku
 
-**LanguageModelLayers** (`src/amber/language_model/layers.py`):
+**LanguageModelLayers** (`src/mi_crow/language_model/layers.py`):
 - Zarządzanie warstwami i hookami
 - Metody do rejestracji/odrejestracji hooków na warstwach
 
-**LanguageModelActivations** (`src/amber/language_model/activations.py`):
+**LanguageModelActivations** (`src/mi_crow/language_model/activations.py`):
 - Zarządzanie aktywacjami modelu
 - Dostęp do zapisanych aktywacji z Store
 
-**LanguageModelTokenizer** (`src/amber/language_model/tokenizer.py`):
+**LanguageModelTokenizer** (`src/mi_crow/language_model/tokenizer.py`):
 - Wrapper tokenizera
 - Ujednolicony interfejs do operacji tokenizacji
 
-**InferenceEngine** (`src/amber/language_model/inference.py`):
+**InferenceEngine** (`src/mi_crow/language_model/inference.py`):
 - Silnik inferencji dla LanguageModel
 - Parametry metody `execute_inference()`:
   - `texts`: `Sequence[str]` - sekwencja tekstów wejściowych
@@ -204,13 +204,13 @@
   - `autocast_dtype`: `torch.dtype | None` - opcjonalny dtype dla autocast
   - `with_controllers`: `bool` - użycie controllerów podczas inferencji (domyślnie `True`)
 
-**LanguageModelContext** (`src/amber/language_model/context.py`):
+**LanguageModelContext** (`src/mi_crow/language_model/context.py`):
 - Dataclass z kontekstem współdzielonym dla LanguageModel i jego komponentów
 - Zawiera: `model`, `tokenizer`, `model_id`, `store`, `device`, `dtype`, rejestry hooków
 
-### 2.4 Moduł amber.mechanistic.sae (moduły uczenia maszynowego)
+### 2.4 Moduł mi_crow.mechanistic.sae (moduły uczenia maszynowego)
 
-**Sae** (`src/amber/mechanistic/sae/sae.py`):
+**Sae** (`src/mi_crow/mechanistic/sae/sae.py`):
 - Klasa abstrakcyjna dziedzicząca z `Controller` i `Detector`
 - Parametry inicjalizacji:
   - `n_latents`: `int` - liczba neuronów latentnych
@@ -225,19 +225,19 @@
   - `modify_activations(module, inputs, output) -> torch.Tensor | None` - modyfikacja aktywacji (Controller)
   - `save(name: str)` - zapis modelu
 
-**TopKSae** (`src/amber/mechanistic/sae/modules/topk_sae.py`):
+**TopKSae** (`src/mi_crow/mechanistic/sae/modules/topk_sae.py`):
 - Implementacja TopK SAE dziedzicząca z `Sae`
 - Dodatkowy parametr: `k: int` - liczba aktywnych neuronów (TopK)
 - Metody:
   - `train(store, run_id, layer_signature, config)` - trening SAE używając aktywacji z Store
-  - `save(name, path)` - zapis modelu (overcomplete state dict + metadata Amber)
+  - `save(name, path)` - zapis modelu (overcomplete state dict + metadata mi_crow)
   - `load(path)` - statyczna metoda do ładowania TopKSae z pliku
 
-**SaeTrainer** (`src/amber/mechanistic/sae/sae_trainer.py`):
+**SaeTrainer** (`src/mi_crow/mechanistic/sae/sae_trainer.py`):
 - Klasa trenująca SAE używając funkcji z biblioteki `overcomplete`
 - Metoda `train(store, run_id, layer_signature, config)` - główna metoda treningu
 
-**SaeTrainingConfig** (`src/amber/mechanistic/sae/sae_trainer.py`):
+**SaeTrainingConfig** (`src/mi_crow/mechanistic/sae/sae_trainer.py`):
 - Dataclass z konfiguracją treningu SAE
 - Parametry treningu:
   - `epochs`: `int` (domyślnie 1)
@@ -265,7 +265,7 @@
   - `wandb_mode`: `str` - tryb wandb: "online", "offline", "disabled" (domyślnie "online")
   - `wandb_slow_metrics_frequency`: `int` - częstotliwość logowania wolnych metryk (L0, dead features) (domyślnie 50)
 
-**AutoencoderContext** (`src/amber/mechanistic/sae/autoencoder_context.py`):
+**AutoencoderContext** (`src/mi_crow/mechanistic/sae/autoencoder_context.py`):
 - Dataclass z kontekstem SAE
 - Parametry:
   - `autoencoder`: `Sae` - instancja SAE
@@ -285,7 +285,7 @@
   - `bias_init`: `float` (domyślnie 0.0)
   - `init_method`: `str` (domyślnie "kaiming")
 
-**AutoencoderConcepts** (`src/amber/mechanistic/sae/concepts/autoencoder_concepts.py`):
+**AutoencoderConcepts** (`src/mi_crow/mechanistic/sae/concepts/autoencoder_concepts.py`):
 - Zarządzanie konceptami SAE
 - Parametry manipulacji konceptami:
   - `multiplication`: `nn.Parameter` - mnożnik dla każdego neuronu (domyślnie ones)
@@ -295,12 +295,12 @@
   - `disable_text_tracking()` - wyłączenie śledzenia tekstów
 - Atrybut: `dictionary: ConceptDictionary | None` - słownik konceptów
 
-**ConceptDictionary** (`src/amber/mechanistic/sae/concepts/concept_dictionary.py`):
+**ConceptDictionary** (`src/mi_crow/mechanistic/sae/concepts/concept_dictionary.py`):
 - Słownik konceptów z neuronami i ich tekstami
 
-### 2.5 Moduł amber.store
+### 2.5 Moduł mi_crow.store
 
-**Store** (`src/amber/store/store.py`):
+**Store** (`src/mi_crow/store/store.py`):
 - Klasa abstrakcyjna do przechowywania tensorów
 - Parametry inicjalizacji:
   - `base_path`: `Path | str` - bazowa ścieżka katalogu (domyślnie "")
@@ -319,14 +319,14 @@
   - `get_detector_metadata(run_id, batch_index)` - odczyt metadanych detektora
   - `get_detector_metadata_by_layer_by_key(run_id, batch_index, layer, key)` - odczyt konkretnego tensora
 
-**LocalStore** (`src/amber/store/local_store.py`):
+**LocalStore** (`src/mi_crow/store/local_store.py`):
 - Implementacja lokalnego przechowywania dziedzicząca z `Store`
 - Używa `safetensors.torch` do zapisu tensorów
 - Metadane zapisywane jako pliki JSON
 - Tensory zapisywane jako pliki safetensors
 - Struktura katalogów: `base_path/runs_prefix/run_id/batch_index/layer_signature/`
 
-**StoreDataloader** (`src/amber/store/store_dataloader.py`):
+**StoreDataloader** (`src/mi_crow/store/store_dataloader.py`):
 - DataLoader-like klasa do iteracji po aktywacjach w Store
 - Parametry:
   - `store`: `Store` - instancja Store

{mi_crow-0.1.1.post12/src/amber → mi_crow-0.1.1.post14/src/mi_crow}/__init__.py

@@ -1,4 +1,4 @@
-"""Amber: helper package for the Engineer Thesis project.
+"""mi_crow: helper package for the Engineer Thesis project.
 
 This module is intentionally minimal. It exists to define the top-level package
 and to enable code coverage to include the package. Importing it should succeed
@@ -6,7 +6,7 @@ without side effects.
 """
 
 # A tiny bit of executable code to make the package measurable by coverage.
-PACKAGE_NAME = "amber"
+PACKAGE_NAME = "mi_crow"
 __version__ = "0.0.0"
 
 

mi_crow-0.1.1.post14/src/mi_crow/datasets/__init__.py

@@ -0,0 +1,11 @@
+from mi_crow.datasets.base_dataset import BaseDataset
+from mi_crow.datasets.text_dataset import TextDataset
+from mi_crow.datasets.classification_dataset import ClassificationDataset
+from mi_crow.datasets.loading_strategy import LoadingStrategy
+
+__all__ = [
+    "BaseDataset",
+    "TextDataset",
+    "ClassificationDataset",
+    "LoadingStrategy",
+]

{mi_crow-0.1.1.post12/src/amber → mi_crow-0.1.1.post14/src/mi_crow}/datasets/base_dataset.py

@@ -9,8 +9,8 @@ from typing import Any, Dict, Iterator, List, Optional, Union
 
 from datasets import Dataset, IterableDataset, load_dataset, load_from_disk
 
-from amber.datasets.loading_strategy import IndexLike, LoadingStrategy
-from amber.store.store import Store
+from mi_crow.datasets.loading_strategy import IndexLike, LoadingStrategy
+from mi_crow.store.store import Store
 
 
 class BaseDataset(ABC):

{mi_crow-0.1.1.post12/src/amber → mi_crow-0.1.1.post14/src/mi_crow}/datasets/classification_dataset.py

@@ -5,9 +5,9 @@ from typing import Any, Dict, Iterator, List, Optional, Sequence, Union
 
 from datasets import Dataset, IterableDataset, load_dataset
 
-from amber.datasets.base_dataset import BaseDataset
-from amber.datasets.loading_strategy import IndexLike, LoadingStrategy
-from amber.store.store import Store
+from mi_crow.datasets.base_dataset import BaseDataset
+from mi_crow.datasets.loading_strategy import IndexLike, LoadingStrategy
+from mi_crow.store.store import Store
 
 
 class ClassificationDataset(BaseDataset):

{mi_crow-0.1.1.post12/src/amber → mi_crow-0.1.1.post14/src/mi_crow}/datasets/text_dataset.py

@@ -5,9 +5,9 @@ from typing import Any, Dict, Iterator, List, Optional, Sequence, Union
 
 from datasets import Dataset, IterableDataset, load_dataset
 
-from amber.datasets.base_dataset import BaseDataset
-from amber.datasets.loading_strategy import IndexLike, LoadingStrategy
-from amber.store.store import Store
+from mi_crow.datasets.base_dataset import BaseDataset
+from mi_crow.datasets.loading_strategy import IndexLike, LoadingStrategy
+from mi_crow.store.store import Store
 
 
 class TextDataset(BaseDataset):

mi_crow-0.1.1.post14/src/mi_crow/hooks/__init__.py

@@ -0,0 +1,20 @@
+from mi_crow.hooks.hook import Hook, HookType, HookError
+from mi_crow.hooks.detector import Detector
+from mi_crow.hooks.controller import Controller
+from mi_crow.hooks.implementations.layer_activation_detector import LayerActivationDetector
+from mi_crow.hooks.implementations.model_input_detector import ModelInputDetector
+from mi_crow.hooks.implementations.model_output_detector import ModelOutputDetector
+from mi_crow.hooks.implementations.function_controller import FunctionController
+
+__all__ = [
+    "Hook",
+    "HookType",
+    "HookError",
+    "Detector",
+    "Controller",
+    "LayerActivationDetector",
+    "ModelInputDetector",
+    "ModelOutputDetector",
+    "FunctionController",
+]
+

{mi_crow-0.1.1.post12/src/amber → mi_crow-0.1.1.post14/src/mi_crow}/hooks/controller.py

@@ -6,9 +6,9 @@ from typing import TYPE_CHECKING
 import torch
 import torch.nn as nn
 
-from amber.hooks.hook import Hook, HookType, HOOK_FUNCTION_INPUT, HOOK_FUNCTION_OUTPUT
-from amber.hooks.utils import extract_tensor_from_input, extract_tensor_from_output
-from amber.utils import get_logger
+from mi_crow.hooks.hook import Hook, HookType, HOOK_FUNCTION_INPUT, HOOK_FUNCTION_OUTPUT
+from mi_crow.hooks.utils import extract_tensor_from_input, extract_tensor_from_output
+from mi_crow.utils import get_logger
 
 if TYPE_CHECKING:
     pass