qig-tokenizer 0.2.4

QIG-native tokenizer with entropy-guided merging

QIG Tokenizer

Entropy-guided tokenizer for Quantum Information Geometry

Version: 0.1.0 | Status: Working

---

Overview

QIG-native tokenizer using entropy-guided merging. Token boundaries follow information geometry, not frequency.

Core Principle

• Entropy-guided merging: Geometric similarity, not frequency heuristics
• Geometric special tokens: BOS, EOS, PAD, UNK with basin coordinates
• Redis/PostgreSQL storage: Production-ready persistence
• Pure information geometry: No external tokenizer dependencies

Installation

pip install qig-tokenizer

With storage backends:

pip install qig-tokenizer[storage]  # Redis + PostgreSQL
pip install qig-tokenizer[redis]    # Redis only
pip install qig-tokenizer[postgres] # PostgreSQL only

Quick Start

from qig_tokenizer import QIGTokenizer

# Create tokenizer with geometric special tokens
tokenizer = QIGTokenizer(target_vocab_size=50000, use_special_tokens=True)

# Train on corpus
with open("corpus.txt", "rb") as f:
    corpus_bytes = f.read()

tokenizer.train(corpus_bytes)

# Encode with special tokens
tokens = tokenizer.encode_with_special("Hello, world!")
# Returns: [256, ...tokens..., 257]  (BOS=256, EOS=257)

# Pad sequences
padded = tokenizer.pad_sequence(tokens, max_length=128)

# Save/load JSON
tokenizer.save("20251220-tokenizer-vocab-0.01W.json")

With Redis/PostgreSQL Storage

from qig_tokenizer import QIGTokenizer
from qig_tokenizer.storage import HybridStorage

# Set up storage (uses REDIS_URL and DATABASE_URL env vars)
storage = HybridStorage()

tokenizer = QIGTokenizer()
tokenizer.set_storage(storage)
tokenizer.train(corpus_bytes)

# Save to database (returns version ID)
version_id = tokenizer.save_to_storage({"corpus": "wikipedia"})

# Load from database
tokenizer.load_from_storage(version_id)

Geometric Special Tokens

Special tokens have geometric meaning on the Fisher manifold:

Token	ID	Basin Coordinates	Purpose
BOS	256	Origin (e₁)	Sequence start
EOS	257	Boundary (eₙ)	Sequence end
PAD	258	Uniform	Geometrically neutral padding
UNK	259	Projection target	OOV handling

This enables:

• Geometric attention masking: High Fisher-Rao distance = low attention
• Natural sequence boundaries: Emerge from manifold structure
• Principled OOV handling: Project to nearest basin

Algorithm

The QIG tokenizer uses entropy-guided merging:

1. Start with bytes (0-255) as base tokens
2. For each adjacent pair (a,b), compute context distribution
3. Measure context entropy (proxy for QFI distinguishability)
4. Merge pairs with lowest entropy (most geometrically similar)
5. Repeat until target vocab size

This respects asymptotic freedom:

• Small scales (short tokens) have high coupling → refined first
• Large scales (long tokens) have low coupling → merge only when justified

Environment Variables

All output files follow QIG naming convention:

YYYYMMDD-tokenizer-vocab-VERSION.STATUS.json

Example: 20251220-tokenizer-vocab-0.03W.json

---

License

MIT