jevmlx

jevmlx is an innovative library designed for Apple Silicon that enables parallel constrained decision-making for machine learning tasks, specifically tailored for MLX (Machine Learning on Apple). Leveraging a typed and schema-valid JSON approach, jevmlx streamlines decision-making processes in one efficient forward pass, enhancing both accuracy and performance.

Key Features

• Schema-Based Decision Making: Users provide a schema (using Pydantic or raw JSON) along with a context, allowing jevmlx to prefill data and make decisions across multiple fields in a single batched operation—no text generation and no need for JSON repair.

• Validated Output with Probabilities: Each decision results in a validated typed object alongside a confidence score for each field, ensuring informed decision-making.

How it Works

jevmlx follows a structured process:

1. The schema is compiled into a decision plan.
2. A prefill operation occurs only once, followed by a batched pass that evaluates all requested fields efficiently.
3. Decision probabilities for each field are calculated directly from the model's predictions, eliminating the chance of malformed outputs.

Illustrative Command-Line Usage:

jevmlx decide --preset fintech_fraud

This command executes a predefined decision-making task concerning fintech fraud.

Python Integration

Using jevmlx in Python consists of defining schemas with Pydantic or standard dictionaries:

from typing import Literal
from pydantic import BaseModel, Field
import jevmlx

class Fraud(BaseModel):
    is_fraudulent: bool = Field(description="Whether the transaction is fraudulent")
    risk_tier: Literal["LOW", "MEDIUM", "HIGH", "CRITICAL"] = Field(description="Risk tier")

context = "Wire transfer to a new IBAN, requested from a Tor exit node on an unrecognized device"
d = jevmlx.decide(Fraud, context, model="mlx-community/Qwen2.5-1.5B-Instruct-4bit")

This example demonstrates how to define a schema and make decisions programmatically.

Benefits Over Traditional JSON Generation

Rather than merely generating JSON outputs, jevmlx provides:

• Always-Valid JSON Objects: Assembling results from pre-decided fields ensures the output adheres to structural requirements and avoids common errors associated with JSON generation.
• Real-Time Confidence Assessment: Users retrieve actionable insights from per-field probabilities, empowering smarter choices based on model predictions.

Evaluation and Benchmarking

jevmlx includes tools for evaluating performance through benchmark tests, offering insights into accuracy and latency, essential for data-driven applications.

Planned Features

Future developments for jevmlx include enhancements in model compatibility and additional user-friendly features to facilitate integration into existing workflows.

For more details, tutorials, and specifications, please refer to the repository's README and documentation.