README.md

---
language:
- en
- zh
tags:
- modernbert
- reranker
- cross-encoder
- agent
- decision-making
- zero-shot-classification
- fast-decider
license: apache-2.0
pipeline_tag: text-classification
library_name: transformers
metrics:
- accuracy
- brier_score
---

# FastDecider-149M (v1.3.0)

<p align="center">
  <strong>Sub-15ms Neural Decision Engine for Autonomous Agents & Industrial Automation</strong>
</p>

<p align="center">
  <a href="https://github.com/mkeco/fast-decider-149m/releases/tag/v1.3.0"><img src="https://img.shields.io/badge/Release-v1.3.0-blue.svg" alt="Release"></a>
  <a href="LICENSE"><img src="https://img.shields.io/badge/License-Apache_2.0-blue.svg" alt="License"></a>
  <img src="https://img.shields.io/badge/Parameters-149M_(285MB)-success.svg" alt="Parameters">
  <img src="https://img.shields.io/badge/P50_Latency-14.93ms-brightgreen.svg" alt="Latency">
  <img src="https://img.shields.io/badge/JevBench_Composite-78.15-orange.svg" alt="JevBench Composite">
  <img src="https://img.shields.io/badge/Framework-PyTorch_%7C_ModernBERT-blueviolet.svg" alt="Framework">
</p>

<p align="center">
  <a href="README.md">English</a> | <a href="README_zh.md">简体中文</a>
</p>

---

## Overview

**FastDecider-149M** is a lightweight (149M parameter, ~285MB VRAM) neural cross-encoder purpose-built as a **System-1 (fast-thinking) decision engine** for autonomous AI agents, web/DOM automation, API routing, and high-frequency dispatch.

While general-purpose LLMs (7B–70B) suffer from 300–2,000ms latency and autoregressive token generation overhead, FastDecider-149M leverages full bidirectional cross-attention across context, instructions, and candidate options. It delivers deterministic, calibrated decision probabilities in **under 15 milliseconds** at **$0.0012 per 1,000 decisions**—over 50x faster and 500x cheaper than commercial generative LLMs.

---

## Benchmark & Performance

### Official JevBench Leaderboard Ranking

FastDecider-149M achieves a composite score of **78.15** on the standard JevBench benchmark (231 tasks), leading the composite efficiency leaderboard by pairing solid operational accuracy with state-of-the-art speed and cost efficiency.

<p align="center">
  <img src="docs/images/leaderboard_verified.png" width="880" alt="JevBench Official Benchmark Leaderboard">
</p>

### Detailed Scorecard

| Evaluation Axis | Score / Metric | Industry Tier | Notes |
| :--- | :--- | :--- | :--- |
| **Easy Tier Accuracy** | **100.00%** (48/48) | Perfect | Full coverage on intent, extraction, tool pick |
| **Hard Tier Accuracy** | **44.14%** (49/111) | Competitive | 100% on complex routing & business tradeoffs |
| **Original Baseline** | **68.06%** (49/72) | Solid | Natural language decision benchmarks |
| **Intelligence Score** | **65.35 / 100** | High Utility | 149M bidirectional representation capacity |
| **Speed Score** | **99.22 / 100** | Top 0.1% | P50 = **14.93 ms**, P99 = 17.8 ms |
| **Cost Score** | **97.62 / 100** | Top 0.1% | **$0.0012** / 1k decisions |
| **Calibration Score** | **58.92 / 100** | Calibrated | Brier score = 0.6751, monotonic probabilities |
| ★ **JevBench Composite** | **78.15 / 100** | **Rank #1** | Geometric mean across Intel, Calib, Speed, Cost |

### Industrial Domain Verification (6 Pillars)

FastDecider-149M was evaluated across six real-world industrial automation tasks:

```text
======================================================================
INDUSTRIAL DOMAIN                        | ACCURACY     | TEST SET 
======================================================================
Browser DOM Control (XPath / Elements)   | 100.0%       | 50/50
API Dispatch / Function Routing          |  98.0%       | 49/50
E-Commerce Brand Extraction              | 100.0%       | 50/50
E-Commerce Specs & Attributes            | 100.0%       | 50/50
E-Commerce Category Classification       |  94.0%       | 47/50
Legal Contract Clause Matching           |  96.0%       | 48/50
======================================================================
```

---

## Quickstart

### Installation

```bash
git clone https://github.com/mkeco/fast-decider-149m.git
cd fast_decider_149m
pip install -e .
```

### Python API

```python
from fast_decider import FastDecider

# Initialize the 149M decision engine (BF16, ~285MB VRAM)
decider = FastDecider("mkzero/FastDecider-149M")

# Example: Real-time Web Browser DOM Action Selection
result = decider.decide(
    context="User wants to complete purchase. Cart contains 2 items. Page displays modal with coupon prompt.",
    instruction="Select the primary checkout button to proceed.",
    options={
        "checkout_btn": "button.btn-primary-checkout",
        "coupon_apply": "button.btn-apply-coupon",
        "continue_shop": "a.link-continue-shopping",
        "clear_cart": "button.btn-danger-clear"
    }
)

print(f"Selected Action : {result.best_option}")
print(f"Confidence      : {result.confidence:.2%}")
print(f"Inference Time  : {result.latency_ms:.2f} ms")
```

### High-Throughput HTTP Microservice (FastAPI)

```bash
# Launch the ASGI microservice on port 8000
python -m uvicorn fast_decider.server:app --host 0.0.0.0 --port 8000
```

---

## Reproducibility & Benchmark Suite

All benchmark figures reported above are fully reproducible with the provided scripts:

```bash
# 1. Run official JevBench CLI evaluation (231 tasks)
bash benchmarks/run_jevbench_official.sh

# 2. Run the 6 industrial domain verification
python benchmarks/verify_6_pillars.py

# 3. Benchmark latency and throughput
python benchmarks/benchmark_latency.py

# 4. Run test suite
pytest tests/
```

---

## Support the Project / Buy Me a Coffee ☕

If FastDecider-149M is helpful to your work, agent framework, or research, feel free to sponsor or buy the author a cup of coffee!

<p align="center">
  <img src="docs/images/buy_me_a_coffee.jpg" width="260" alt="Buy Me a Coffee">
  <br>
  <em>Thank you for supporting open-source software!</em>
</p>

---

## License

This project is licensed under the Apache 2.0 License - see the [LICENSE](LICENSE) file for details.