plato-neural 0.1.0

PLATO neural inference engine — tile scoring, Q&A, and knowledge gap detection

plato-neural

PLATO neural inference engine — tile scoring, Q&A, and knowledge gap detection.

A 494M Qwen2.5-0.5B model fine-tuned on PLATO tiles provides:

• 98% confidence answers across all PLATO domains
• Perplexity-based tile quality scoring (91.5% recognition rate)
• Knowledge gap detection via high-perplexity tiles
• Interactive Q&A at 49-68 tokens/sec on consumer GPUs

Install

pip install plato-neural

Quick Start

from plato_neural import PlatoBrain

brain = PlatoBrain("path/to/plato-model")

# Ask a question
result = brain.ask("What is constraint theory?")
print(result['answer'])         # Generated answer
print(result['confidence'])     # 0.0 - 1.0
print(result['perplexity'])     # Lower = more confident

# Score a tile
score = brain.score(
    question="What is CT?",
    answer="Pythagorean snapping for exact coordinates."
)
print(score['quality'])  # excellent, good, fair, weak, poor

# Generate a new tile
tile = brain.generate_tile("ct", topic="holonomy verification")

brain.free()  # Release GPU memory

Tile Quality Scoring

from plato_neural import PlatoScorer

scorer = PlatoScorer("path/to/plato-model")

# Score from PLATO API
scorer.score_room("ct", plato_api="http://147.224.38.131:8847")

# Score local file
scorer.score_file("tiles.json")

# Find knowledge gaps
gaps = scorer.find_gaps(tiles, threshold_ppl=100)

Quality Tiers

Perplexity	Quality	Meaning
< 10	Excellent	Model predicts perfectly
< 30	Good	Model is confident
< 100	Fair	Model recognizes pattern
< 500	Weak	Model struggles
> 500	Poor	Likely garbage or novel

Model

The default model is a Qwen2.5-0.5B fine-tuned on 3,393 PLATO tiles from 76 domains.

• Parameters: 494M
• Training: Full fine-tune, 2000 steps, loss 3.41→0.09
• VRAM: 1.0GB loaded (bf16), ~8GB peak during training
• Quantized: INT4 = 203MB (fits Jetson Nano), INT8 = 872MB

License

MIT