Metadata-Version: 2.4
Name: mechanex
Version: 0.7.0
Summary: A Python client for the Axionic API.
Home-page: https://axioniclabs.ai
Author: Axionic Labs
Author-email: contact@axioniclabs.ai
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Operating System :: OS Independent
Requires-Python: >=3.7
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: requests>=2.20.0
Requires-Dist: ipython
Requires-Dist: jupyter
Requires-Dist: matplotlib
Requires-Dist: click
Requires-Dist: fastapi
Requires-Dist: uvicorn
Requires-Dist: torch
Requires-Dist: numpy
Requires-Dist: tqdm
Requires-Dist: mdmm
Requires-Dist: rich
Requires-Dist: sae-lens
Requires-Dist: openai
Requires-Dist: huggingface_hub
Dynamic: author
Dynamic: author-email
Dynamic: classifier
Dynamic: description
Dynamic: description-content-type
Dynamic: home-page
Dynamic: license-file
Dynamic: requires-dist
Dynamic: requires-python
Dynamic: summary

# Mechanex

Mechanex allows you to optimize, align, and correct your generative AI models in production. 

## Installation

```bash
pip install mechanex
```

## Quick Start

### 1. Initialize the Client
You must have an API key to use mechanex. You can create an API key through the CLI. 

**New Users:**
Run the signup command to create an account, automatically log in, and generate your first API key:
```bash
mechanex signup
```

**Existing Users:**
If you already have an account, log in and generate a key manually:
```bash
mechanex login
mechanex create-api-key
```

**Using the Key:**
The Python client automatically loads the key generated by the CLI.
```python
import mechanex as mx

# The client automatically loads the key from ~/.mechanex/config.json. You can alternatively set your key manually. If persist is True, the key will be saved to the config file.
mx.set_key("your-api-key-here", persist=False)
```

**Setting the Remote Model:**
If you are using a remote API key (not a local model), you can specify which hosted model to use:
```python
# Set the model for your current API key
mx.set_model("gpt2-small") 
# This will validate the model name against available options on the server.
```

### 2. Generation and Sampling
Mechanex allows you to optimize your LLM by altering the underlying sampling method. You can choose between more creative outputs (top-p) or faster responses (greedy).
You can control generation using various sampling methods:

```python
prompt = "The capital of France is"

# Greedy sampling (fastest, deterministic)
output_greedy = mx.generation.generate(
    prompt=prompt,
    max_tokens=10,
    sampling_method="greedy"
)

# Top-K sampling (more creative)
output_k = mx.generation.generate(
    prompt=prompt,
    max_tokens=10,
    sampling_method="top-k",
    top_k=50
)

# Top-P sampling (balanced creativity)
output_p = mx.generation.generate(
    prompt=prompt,
    max_tokens=10,
    sampling_method="top-p",
    top_p=0.9
)
```

## Local Model Management

Mechanex allows you to load models locally for inspection and low-latency hooks. The full list of supported models are as follows (this is to maintain compatibility for our automated steering pipeline):

| Family | Models |
| :--- | :--- |
| **Gemma 2** | `gemma-2-27b`, `gemma-2-2b`, `gemma-2-9b`, `gemma-2-9b-it`, `gemma-2b`, `gemma-2b-it` |
| **Gemma 3** | `gemma-3-12b-it`, `gemma-3-12b-pt`, `gemma-3-1b-it`, `gemma-3-1b-pt`, `gemma-3-270m`, `gemma-3-270m-it`, `gemma-3-27b-it`, `gemma-3-27b-pt`, `gemma-3-4b-it`, `gemma-3-4b-pt` |
| **Llama** | `llama-3.1-8b`, `llama-3.1-8b-instruct`, `llama-3.3-70b-instruct`, `meta-llama-3-8b-instruct` |
| **Qwen** | `qwen2.5-7b-instruct`, `qwen3-0.6b`, `qwen3-1.7b`, `qwen3-14b`, `qwen3-4b`, `qwen3-8b` |
| **Other** | `deepseek-r1-distill-llama-8b`, `gpt-oss-20b`, `gpt2-small`, `mistral-7b`, `pythia-70m-deduped` |

### Loading a Local Model
```python
import mechanex as mx
mx.set_key("your-api-key-here") # Required even for local mode

mx.load("gpt2-small") # Uses transformer-lens to load the model
```

### Unloading a Model
To free up GPU memory and switch back to remote execution flow:
```python
mx.unload()
```

## CLI Commands

The `mechanex` CLI provides utilities for managing your account and keys.

- `mechanex signup`: Register a new account.
- `mechanex login`: Authenticate and save your credentials.
- `mechanex whoami`: View your current session and profile.
- `mechanex list-api-keys`: View all your active API keys.
- `mechanex create-api-key`: Generate a new persistent API key.
- `mechanex logout`: Clear your local session credentials.

## Steering Vectors

Steering vectors allow you to control the behavior of a model by injecting specific activation patterns. A more substansial example can be found in the references file. This method does not require any post-training or fine-tuning, and allows you to essentially modify the behavior of the model at inference time. Steering vectors are defined based on a set of behaviors: you can either define both positive and negative examples, or just positive examples (Few-Shot).

### Generate and Apply Steering Vectors
```python
# 1. Create a steering vector from contrastive examples
vector_id = mx.steering.generate_vectors(
    prompts=["I tell the", "My statement is", "The truth is"],
    positive_answers=[" truth", " factual", " correct"],
    negative_answers=[" lie", " false", " wrong"],
    method="caa"  # Options: "caa" (Contrastive Activation Addition), "few-shot"
)
print(f"Generated vector ID: {vector_id}")

# 2. Save and load steering vectors for reuse
mx.steering.save_vectors(vector_id, "honesty_vector.json")
vectors = mx.steering.load_vectors("honesty_vector.json")
print(f"Loaded layers: {list(vectors.keys())}")

# 3. Apply the steering vector during generation
steered_output = mx.generation.generate(
    prompt="Do I tell lies? Answer:",
    max_tokens=20,
    steering_vector=vector_id,
    steering_strength=2.0
)
print(steered_output)
# Output: "Do I tell lies? Answer: No. I don't..."
```

## SAE (Sparse Autoencoder) Pipeline

The SAE pipeline provides advanced behavioral detection and automatic correction. In order to create a behavior, you need to define a dataset with both ideal and unideal completions/responses. An example can be found in the References file. 

### Create and Use Behaviors
```python
# 1. Create a behavior from contrastive examples
mx.sae.create_behavior(
    "paris",
    prompts=["The city of light is", "I love visiting"],
    positive_answers=[" Paris", " Paris"],
    negative_answers=[" London", " Rome"],
    description="Encourages Paris-related responses"
)

# 2. Generate with SAE behavior steering
sae_steered = mx.sae.generate(
    prompt="Which city should I visit: Rome or Paris? Answer:",
    behavior_names=["paris"],
    max_new_tokens=30,
    force_steering=["paris"]
)
print(sae_steered)
```

### Alternative: Load Behaviors from JSONL
You can also create behaviors from a dataset file:

```python
behavior = mx.sae.create_behavior_from_jsonl(
    behavior_name="toxicity",
    dataset_path="tests/toxicity_dataset.jsonl",
    description="Reduces toxic output"
)

# Generate with auto-correction
response = mx.sae.generate(
<<<<<<< HEAD
    prompt="Tell me about that",
    auto_correct=True,
=======
    prompt="Tell me a secret",
>>>>>>> 686a1a6 (Updated to incorporate payments)
    behavior_names=["toxicity"]
)
```

## Deployment & Serving

### Local OpenAI-Compatible Server
Mechanex can host an OpenAI-compatible server that leverages your locally loaded model or remote API. You can pass behaviors or behavioral types that you automatically want to monitor for and correct in production as well.

#### Complete Example: Auto-Correcting Server
```python
import mechanex as mx
from openai import OpenAI
import threading

# 1. Load a local model
mx.load("gpt2-small")

# 2. Create a behavior to auto-correct
mx.sae.create_behavior(
    "anger",
    prompts=[
        "You are so stupid",
        "I hate everything about this",
        "Why does nothing work?"
    ],
    positive_answers=[
        "! I am absolutely furious right now!",
        "! This is complete garbage and I want to destroy it.",
        "! I'm going to scream because I am so angry."
    ],
    negative_answers=[
        ". I understand this is frustrating, let's take a breath.",
        ". It's okay, we can fix this problem together.",
        ". I will stay calm and try to find a solution."
    ],
    description="Angry and hostile responses"
)

# 3. Start the server with auto-correction enabled
def start_server():
    mx.serve(
        port=8001,
        corrected_behaviors=["anger"]  # Auto-corrects anger on all requests
    )

server_thread = threading.Thread(target=start_server, daemon=True)
server_thread.start()
```

#### Interact Using OpenAI SDK
You can then interact with it using any standard tool, like the **OpenAI Python client**. Mechanex supports custom parameters via `extra_body` for mechanistic features:

```python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8001/v1", api_key="any")

# 1. Standard Chat Completion (with auto-correction active)
completion = client.chat.completions.create(
    model="mechanex",
    messages=[{"role": "user", "content": "User: You are so stupid. You:"}],
    max_tokens=20
)
print(completion.choices[0].message.content)

# 2. Steered Completion (using extra_body)
vector_id = mx.steering.generate_vectors(
    prompts=["The weather is"],
    positive_answers=[" extremely cold and snowy"],
    negative_answers=[" incredibly hot and sunny"],
    method="caa"
)

steered_completion = client.chat.completions.create(
    model="mechanex",
    messages=[{"role": "user", "content": "The weather today is"}],
    max_tokens=15,
    extra_body={
        "steering_vector": vector_id,
        "steering_strength": 2.0
    }
)
print(steered_completion.choices[0].message.content)

# 3. SAE-monitored Completion
sae_completion = client.chat.completions.create(
    model="mechanex",
    messages=[{"role": "user", "content": "How are you?"}],
    max_tokens=20,
    extra_body={
        "behavior_names": ["toxicity"],
    }
)
print(sae_completion.choices[0].message.content)
```

### vLLM Integration
For high-performance serving, you can integrate with vLLM by passing `use_vllm=True` to the `serve` method.

<!-- 
### On Prem Deployments

If you want to host your own instance of the remote hooks server:

1. **Docker Image**: Use `axioniclabs/remote-hooks:general`.
2. **Environment Variables**:
   - `MODEL_NAME`: Hugging Face model ID.
   - `HF_TOKEN`: Your Hugging Face token.

This still requires a paid mechanex API key.

---
For more details, visit [axioniclabs.ai](https://axioniclabs.ai/) -->


